Continual RL

In Continual Multi-Agent Reinforcement Learning (CMARL), agents learn cooperative policies across sequences of tasks, aiming to adapt effectively to new tasks while preserving the ability to solve previously encountered ones. In many applications, tasks differ in their underlying structure, which can represent, for example, distinct operational conditions or target configurations (e.g., different network topologies in power grids or arrangements in formation control). Existing CMARL methods lack dedicated mechanisms to leverage this structural information when learning new tasks, failing to promote transfer and mitigate forgetting. To fill this gap, we propose Continual Graph Multi-Agent Reinforcement Learning (CGMARL), a novel framework for CMARL problems in which task sequences are mapped into a series of attributed graphs, each modeling a task-specific structure. In CGMARL, each graph determines the environment dynamics (next states and/or rewards) and the number of agents for the corresponding task. Then, we present Graph-based Formation (GRAFO), the first CGMARL benchmark, and show how forgetting arises in this setting. Finally, to address this limitation, we propose Frozen Graph Encoder (FROG), a method that relies on a frozen graph backbone to preserve past structural information in graph-based CMARL policies. Experiments on GRAFO show that pairing FROG with existing CL methods substantially improves performance on multiple CGMARL scenarios.

A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wisdom is that on-policy training is a prerequisite for continual learning. In practice, however, data containing new knowledge or capabilities are often off-policy. While methods such as on-policy self-distillation (OPSD) try to bridge this gap by converting off-policy data into on-policy signal, they have been shown to cause reasoning collapse. In this paper, we show that off-policy merging beats OPSD for continual learning. We first show that SFT learns a useful signal from new data, but naively applying its update interferes with existing capabilities. We reduce this interference with a simple recipe we term grafting, which changes where the update is learned and how it is applied: (1) learning the update on an earlier donor checkpoint, ideally even before the end of pretraining, and applying the weight update to the post-trained model; (2) scaling the weight update, equivalent to a form of model merging; and (3) optionally, masking the most sensitive update directions when the new data distribution is far from the post-trained model. Across continual learning settings including (1) distilling from expert traces, (2) self-improvement with STaR and Pedagogical RL, and (3) injecting knowledge after pretraining cutoff, grafting Pareto-dominates both SFT and OPSD in new-task and old-task performance, while avoiding expensive on-policy sampling. Therefore, our work challenges on-policy training as a necessity for continual learning on RL-trained models.

Despite many studies about causes and remedies of plasticity loss in Reinforcement Learning (RL) under continual task changes, no RL method has yet consistently achieved a good balance between adaptation and forgetting. Here we turn to an alternative optimization paradigm, neuroevolution (NE): algorithms that search directly in weight space through mutation and selection over a population of neural networks. Across a wide array of environments and environmental changes, with policies ranging from a few hundred parameters to million-parameter networks, we compare evolution strategies (ES) and genetic algorithms (GAs) against state-of-the-art continual RL variants and population-based RL. ES most consistently achieves a good stability-plasticity trade-off, while the GA is the most plastic method but forgets more than ES. To explain this, we study the return landscape around each method's solutions. ES finds the widest neighborhoods, i.e.\ regions of weight space in which perturbed policies still solve the task, and the size of the overlap between the neighborhoods of consecutive tasks correlates with a method's stability-plasticity trade-off. Rewarding behavioral diversity in a GA through novelty search makes the population even more plastic, at the cost of forgetting. Finally, symptoms of plasticity loss commonly reported in RL do not transfer to NE. Overall, these results establish NE as a competitive alternative to RL under continual task changes, and suggest that training under perturbations in weight space may be a useful mechanism for continual learning more broadly.

Independent evaluation can reject harmful policy updates yet also prevent useful continual learning. We argue that update admission must be assessed through both error control and retained learning opportunities at a stated interaction budget. We identify a concrete failure: a range-based confidence gate cannot certify unchanged old-task behavior within otherwise substantial budgets. A standard paired-binomial construction reduces this burden when outcome disagreements are rare. We also specify certified historical-reference promotion and a round-level missed-opportunity metric. In a constructed one-step pushing diagnostic with 32 seeds, fresh paired checks admit 31.6% of a common update stream at 2,000 episodes per stage, versus zero for the range-based gate; unconditional replay nevertheless learns better in closed-loop runs. A separate learned-dynamics stress test distinguishes model bias from feedback-selection error. The contribution is an admission-audit protocol with analytical and synthetic evidence; physical-robot and VLA validation remain open.

In continual reinforcement learning, carefully managing the stability-plasticity tradeoff remains a core challenge. Recent work by Abel et al. (2025) formalized this dilemma by defining plasticity as the generalized directed information from an agent's observations to its actions, and empowerment as the generalized directed information from its actions to its observations. This formulation successfully reframes the traditional stability-plasticity tradeoff as an empowerment-plasticity tradeoff. However, while extensive literature exists on optimizing for empowerment, there is currently no research addressing the optimization of plasticity under this new definition. This paper presents preliminary work toward optimizing plasticity within Markov decision processes. We show that there exists a Bellman optimality equation for optimizing plasticity similar to previous work for empowerment.

Continual world models must decide whether new data justify changing the model. Fixed replay schedules and prediction-error triggers specify when to update, but neither reveals the value of an individual update: one deployment run cannot show how the same model would have performed at that moment had it held its parameters. We introduce the fork ledger, which branches a deployment stream at pre-registered decision points into matched update and hold continuations under common random numbers. It evaluates both continuations on the same episodes and records $\Delta R = R_{\mathrm{update}} - R_{\mathrm{hold}}$. Always applying one fixed update mechanism lowers return on all three simulated control tasks: CartPole ($-144.0$; checkpoint-bootstrap $95\%$ CI $[-185.4,-116.1]$, against a converged return near $650$), Walker ($-82.8$; $[-101.1,-61.7]$) and Cheetah ($-18.6$; $[-29.0,-6.6]$). Divergence is an outcome of applying the update, so the estimand counts every attempted fork; restricted to the $693$ of $720$ that did not collapse, CartPole and Walker are unchanged in sign ($-113.4$ and $-82.1$) and Cheetah becomes unresolved ($-3.9$; $[-17.5,+13.0]$). The task is the unit of inference: each contributes $240$ attempted forks over five pretrained checkpoints crossed with two drift directions. The ledger makes counterfactual utility observable for a fixed mechanism, allowing triggers to be judged by the updates they select rather than by surprise detection alone.

Reinforcement learning with verifiable rewards (RLVR) commonly post-trains reasoning models on multiple tasks, while rerunning multitask RLVR (MTRL) as new tasks are added makes capability expansion costly. We therefore study continual RLVR, which updates the existing model as each task arrives. The central question is whether a model updated this way can perform as well as a jointly trained model. To answer this question, we introduce Continual Reasoning Gym, a continual-RLVR environment that organizes text and visual reasoning tasks into five task sequences. In this setting, we identify two key observations: Sequential RLVR exhibits modest forgetting, yet its final performance remains below that of MTRL. To understand the latter, we decompose final performance and show that forgetting accounts for only part of the gap. To explain the former, we identify shared reasoning: transferable reasoning structure allows training on one task to support others on average. We therefore introduce Continual Prompt Replay (CPR), which harnesses shared reasoning to improve learning on the arriving and future tasks by replaying previous-task prompts and regenerating their responses with the current policy. On average, only CPR reaches MTRL-level performance.

This work explores the relationship between task similarity and catastrophic forgetting in reinforcement learning. Catastrophic forgetting, the phenomenon in machine learning of losing the ability to effectively perform on previous tasks, is a significant impediment to continual learning. This study aims to understand the extent to which the similarity of a new task influences the performance on the previous task. Interpretable reinforcement learning, specifically Q-learning, is employed on graph-based tasks with the objective of minimising the number of steps to reach a goal. The study investigates the performance on a previously learned task after training on a new task, for tasks of varying relative levels of complexity. The experimental results reveal a complex dynamic between task similarity and forgetting, with significant fluctuations in forgetting severity observed across degrees of task similarities and task complexities, and are suggestive of an interdependence of forgetting on the similarity and complexity of tasks. The observations were accompanied by observations of high degrees of variability in forgetting and an uneven distribution of task similarity measures. The relationship between these variables remains unclear and no evidence of statistical significance that task similarity has an effect, independently, on forgetting is found in continual reinforcement learning. Further research is warranted to gain a comprehensive understanding of the potential interplay between task similarity and catastrophic forgetting.

Contemporary model-free reinforcement learning algorithms can achieve very high performance, but have low sample efficiency and are not robust to changes in the environment. Model-based algorithms have much higher sample efficiency, but still fail when the environment shifts. This paper introduces Adaptive Topological Learning with Abstract Successors (ATLAS) to combat these challenges. ATLAS uses a Grow When Required network with Successor Features in order to achieve high sample efficiency while also robustly tackling catastrophic forgetting. We evaluate ATLAS in spatial navigation tasks, benchmarking its performance against common on-policy and off-policy algorithms. Our empirical results demonstrate that by structurally decoupling transition dynamics from the reward signal, ATLAS achieves near-instantaneous adaptation to new goals and can exhibit positive backward transfer, significantly outperforming baseline methods in non-stationary environments.

We show that on-policy reinforcement learning with verifiable rewards (RLVR) can improve the current objective while making successful behaviors for later objectives too rare to sample and reinforce. We call this verifier-induced support reshaping and define effective rewardable support as successful trajectories reachable within a fixed rollout budget. Across two model families, we study this effect through repeated verifier-scored sampling and bidirectional training on mathematical reasoning and constrained instruction following, including sequential training with the opposite verifier. Math-RLVR raises average instruction-following success but reduces the number of prompts with any successful response under repeated sampling. On IFEval with Qwen3-8B-Base, pass@1 rises by 6.5 percentage points while best@32 falls by 9.8 percentage points, and the same divergence appears across both models and IF benchmarks. Conversely, IF-RLVR shifts math responses from step-by-step openings toward direct answers, lowers best@k across sampling budgets, and reduces reward variation for later Math-RLVR. Token-distribution analyses and controlled opening interventions show that these changes concentrate in the first few response tokens. RLVR mainly reranks openings already available in the base policy, and the selected opening causally affects math searchability. The tested reference-policy constraints, routing priors, and on-policy distillation preserve cross-task support only partially; MathIF and ReasonIF show that marginal gains translate only partly into responses that are both correct and constraint-following. Therefore, endpoint improvements do not guarantee future trainability or joint capability under on-policy optimization. Code is available at https://github.com/sylvain-wei/VISR

Neural networks are hindered by accumulating dormant neurons and loss of expressivity throughout training, particularly in non-stationary data settings, such as continual supervised and reinforcement learning. Recently, neuron resets have been used to maintain gradient flow and restore plasticity. However, full unit reinitialization often sacrifices peak performance and can destabilize training, leading to policy collapse. To preserve plasticity without destabilizing training, we propose Calibrated Partial Resets (CPR), an optimizer that periodically pulls low-utility neurons toward their initialization, with pull strength scaled by each neuron's utility. Unlike binary reset methods, partial resets avoid brittleness; unlike uniform decay, calibrated utility-scaling concentrates adjustment on the units that need it most. Among compared methods, only CPR avoids policy collapse over 400M training steps in SlipperyAnt, and it outperforms prior decay and reset-based methods on Continual MetaWorld and Continual MinAtar benchmarks. Ablations reveal a tunable trade-off between plasticity and peak performance, highlighting utility-scaled reinitialization as a promising direction for continual learning.

A key challenge in modern robotics is to adapt to changing environments, a challenge that is exacerbated when simulations cannot encompass every possible real-world configuration, and therefore Reinforcement Learning (RL) in the physical world becomes necessary. Continual Reinforcement Learning provides the tools to address this challenge; however, both the frameworks and the methods remain underexplored. Autonomous Racing and in particular the RoboRacer competition provide a testing ground for such methods, as learning to drive on a new track-floor combination with the least amount of new experience naturally frames a continual learning problem. This work tries to address this gap by proposing a continual RL framework based on Continual Backpropagation that is able, with only real-world data, to train a generalistic policy on a set of tracks and then fine- tune it within 15 minutes to outperform classical controllers. Furthermore, a comparison method based on offline RL is proposed, and a simulation analysis of the plasticity properties of the methods is conducted.

Artificial Intelligence (AI) systems often perform well on isolated tasks but struggle under continual learning conditions, where training on new tasks can overwrite previously acquired knowledge, a failure mode known as catastrophic forgetting. Biological learning systems reduce this interference through complementary memory processes involving rapid hippocampal encoding and slower cortical consolidation. This study introduces NeuroSynth, a brain-inspired continual reinforcement learning architecture designed to mitigate catastrophic forgetting through a dual-pathway consolidation mechanism. NeuroSynth separates rapid task acquisition from long-term retention using distinct "plan" and "habit" pathways combined with replay and knowledge distillation. NeuroSynth was evaluated against Proximal Policy Optimization (PPO) and Elastic Weight Consolidation (EWC) across three sequential navigation tasks with changing goal locations in a non-revisitation continual learning setting. Across six independent seeds, NeuroSynth preserved substantially more early-task knowledge than PPO after sequential training, achieving 18.00% Task A success rate compared to 0.33% for PPO (p = 0.014929, Cohen's d = 1.49) and 35.33% Task B success rate compared to 0.00% for PPO (p = 0.002376, Cohen's d = 2.31). NeuroSynth also demonstrated higher final Task C performance than EWC, achieving 9.00% compared to 2.00% (p = 0.226643, Cohen's d = 0.56), indicating a moderate but not statistically significant advantage. These findings suggest that biologically inspired consolidation mechanisms may improve the stability-plasticity balance in continual reinforcement learning systems.

Effective decision-making in complex and changing environments requires balancing short-term and long-term consequences. In reinforcement learning (RL), this trade-off is typically controlled through a fixed discount factor, which imposes a single exponentially discounted temporal horizon. However, biological agents exhibit flexible and adaptive temporal discounting, suggesting that effective planning requires multiple timescales. Here, we propose a multi-horizon approach that adaptively selects and combines temporal horizons, enabling robust adaptation to changes in reward structure without manual discount-factor tuning. This flexibility makes the method particularly suitable for continual learning scenarios involving task switches and varying environmental configurations. Empirically, we demonstrate that our approach identifies effective discount factors across a range of MiniGrid environments, including continual settings composed of three sequentially changing tasks. These results suggest that adaptive temporal discounting can improve parameter efficiency and enhance adaptability in both artificial and biologically inspired learning systems.

Humans can progressively acquire highly dynamic motor skills while preserving reliable everyday motor abilities. In contrast, existing humanoid controllers face a trade-off between generalist and specialist capabilities: generalist motion tracking policies struggle to reliably execute rare highly dynamic motions, whereas specialist training can degrade previously acquired behaviors. We introduce Extreme-RGMT, a two-stage continual learning framework for robust generalist humanoid control. The method first learns a generalist motion-tracking base policy from diverse multi-source motion data, then employs an asymmetric skill acquisition and capability consolidation mechanism to constrain policy drift on mastered motions while emphasizing difficult dynamic segments. To address the scarcity of highly dynamic motions, their high failure rates, and the resulting shortage of informative samples, Extreme-RGMT combines difficulty-aware sampling with advantage-prioritized trajectory resampling to emphasize critical segments. Experiments show that Extreme-RGMT achieves state-of-the-art generalist whole-body motion-tracking performance, including substantially improved completion of challenging highly dynamic motions. The resulting controller directly executes diverse unseen highly dynamic motions under fixed references and online inertial motion-capture inputs, advancing generalist whole-body motion-tracking controllers toward highly dynamic motor capabilities at the human-expert level.

Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience. We ask a question the continual-RL literature has assumed an answer to but never measured: which component forgets? Under never-clear replay, pre-registered component-level probes (n=3 seeds throughout) show that the world model retains essentially everything measurable about old tasks -- reward discrimination (retention ratio ~1.0), value estimates, and termination structure -- while the actor's behavior collapses. Forgetting in this regime is a channel problem, not a memory problem. We demonstrate this by intervention: with the world model frozen and identical imagined rollouts, reinforcement learning in imagination fails to recover a lost skill (0/3 seeds), while supervised self-imitation on the world model's own graded dreams recovers it on 3/3 seeds with zero environment interaction. Interleaved during training, this graded dream rehearsal yields a task-label-free, parameter-constant continual learner: 3/3 four-task chains retained where plain replay passes 0/3, 3/3 eight-task chains, and consistent gains over matched real-episode cloning (paired difference +0.13, bootstrap 95% CI [0.07, 0.24], complete seed separation). The dream-grading step is load-bearing: we characterize two scoring failure modes, provide an offline selection gauge that caught both before they contaminated results, and give a realized-first grading rule that closes them. All experiments were pre-registered with committed protocols; every refuted hypothesis is reported.

Continual post-training is becoming a central paradigm for adapting vision-language models to evolving tasks. Recent studies report that reinforcement learning is less prone to forgetting than supervised fine-tuning, motivating the view that RL is inherently resistant to forgetting. However, this view remains insufficiently validated, as existing evidence is largely drawn from outdated or homogeneous benchmarks. We revisit this assumption by introducing MRCL, a Multimodal Reasoning Continual Learning benchmark built from recent and diverse multimodal reasoning tasks. Experiments on MRCL show that standard reinforcement learning still suffers from catastrophic forgetting during continual post-training. We trace the failure to an objective mismatch. The KL regularization used in common policy optimization methods is evaluated on current-task data, whereas forgetting is caused by behavioral drift on prior-task distributions. To address this problem, we propose Continual Policy Optimization (CPO), a replay-free method grounded in a prior-task behavioral KL objective. CPO derives a local Fisher surrogate from the historical KL objective and uses parameter movement as a gradient-free proxy for Fisher sensitivity, enabling sparse regularization with negligible additional overhead. Experiments on three model scales and comparisons with multiple RL baselines show that CPO consistently reduces forgetting while maintaining effective adaptation and preserving broader pretrained capabilities. The implementation code is available at https://github.com/MaolinLuo/CPO.

Robot policies inevitably encounter failures when deployed in real environments. Naive retries often repeat the same mistakes, while many existing recovery methods rely on human intervention. In this paper, we propose Failure-Aware Retry (FAR), a framework that enables robots to learn from previous failures at test time, adapt their behavior accordingly, and eventually complete the task autonomously. FAR combines Failure-Contrastive Preference Adaptation, which constructs preference learning data from failures to steer the policy away from previously unsuccessful behaviors, with lightweight action perturbations during retries to encourage local exploration. We further incorporate successful recovery trajectories into a training loop for continual policy improvement. Experiments in both simulation and real-world manipulation tasks show that FAR substantially improves success rates and robustness, with average gains of 17.6% over the standard diffusion policy in simulation and 11.7% in the real world. In addition, FAR improves data efficiency under both reset and timestep budgets during continual policy improvement by exploiting informative failure cases. Videos and code are available at https://hoar012.github.io/FAR-Project.

Autonomous driving policies should be able to improve continually as deployment exposes them to increasingly diverse and long-tail traffic situations. However, most learning-based policies are trained or fine-tuned on expert demonstrations and then rely largely on generalization to handle challenging closed-loop scenarios, lacking an explicit mechanism to correct and retain the mistakes exposed in these scenarios. This paper studies autonomous driving policy improvement from a lifelong learning perspective: Can a pretrained policy improve continually by accumulating corrective knowledge derived from its own mistakes, while retaining previously acquired driving competence? To answer this question, we propose Rollout-Retrieval Lifelong Policy Learning (R$^2$LPL), a policy learning framework that retrieves corrective targets from recoverable policy-induced mistakes and retains the resulting knowledge through lifelong policy learning. R^2LPL addresses a key bottleneck in continual policy improvement: closed-loop mistakes reveal where the policy is weak, but do not directly specify what the policy should learn. By filtering recoverable mistake-related states and retrieving feasible corrective targets, R$^2$LPL turns sparse failure evidence into compact supervised knowledge for stable and sample-efficient policy improvement. We evaluate R$^2$LPL on large-scale closed-loop nuPlan benchmarks. With only a few rollout and continual-learning cycles, R$^2$LPL elevates a learning-based planner with moderate initial performance to state-of-the-art performance across the evaluated benchmarks, especially on the challenging and long-tail Test14-hard split. These results demonstrate the effectiveness of R$^2$LPL in converting recoverable closed-loop mistakes into corrective knowledge for sustained policy improvement.

Robots deployed in the real world rarely operate under a single fixed dynamics model: wind changes, payloads vary, batteries drain, contacts shift, and hardware wears. Yet most learning-based controllers are trained once and deployed as if learning were complete. This prevents the robot from using deployment experience to further improve task performance. In this work, we propose a continual learning framework that uses real-world experience to improve robot policies under hidden and recurring dynamics. Our method learns a condition-aware dynamics model from real state-action trajectories by combining an analytical physics prior with a neural residual for unmodeled effects. A recurrent encoder infers the current hidden condition from recent interaction, and this estimate conditions both the residual model and the policy. Policy learning is performed via differentiable simulation using diverse learned dynamics sampled from the latent model. At deployment, these sampled conditions are replaced by conditions inferred online from recent real interaction, allowing the policy to recover recurring dynamics by recognition rather than residual re-fitting. Through extensive simulation studies and real-world experiments, we demonstrate that the framework improves policy performance under diverse unobserved disturbances. On real quadrotor trajectory tracking under changing wind, the policy recovers from recurring disturbances in roughly 1s, about 5x faster than online residual re-fitting. It also reduces large-disturbance hover and tracking errors by 65.7% and 53.3% over the state-of-the-art online adaptation approaches

Multi-Agent Reinforcement Learning (MARL) value factorization methods can suffer from a loss of plasticity, gradually failing to adapt when transferring to new task instances. We trace this issue to stagnant neurons, units whose gradient updates become negligibly small relative to their weights, thereby hindering learning. While existing plasticity injection methods exist, they prove ineffective for such neurons. To address this, we propose Knowledge-retentive Neuron-level PlastIcity Focusing InjEction (KNIFE), a novel method that directly targets stagnant neurons. KNIFE replaces each stagnant neuron with a composite unit comprising three specialized components: a frozen knowledge neuron to preserve acquired knowledge, a re-initialized active neuron to restore learning capacity, and a compensation neuron to ensure the combined output matches the original, thus maintaining previous learned cooperation knowledge. Extensive experiments on SMACv2, predator-prey, and matrix games demonstrate that KNIFE significantly outperforms state-of-the-art plasticity injection methods.

Self-improvement can self-regress. In REINFORCE post-training for code, a model can quickly improve on its optimized metric and then collapse within the same training campaign. We study this in a controlled multi-seed testbed using Qwen-2.5-3B and Qwen-2.5-7B, trained on competitive-programming tasks with binary CodeGrader reward across 10 sequential 20-step campaigns. Across campaigns, pass@1 shows a robust rise-then-collapse pattern: it peaks within tens of gradient steps and then falls back, sometimes to near zero. This is not cross-task catastrophic forgetting, but within-task policy over-optimization on a fixed distribution; KL- and EWC-style constraints do not prevent it. We ask where the control loop should sit. We compare three levels: CARE, a between-campaign memory mechanism with a capability posterior, transfer gate, and regression-aware belief revision; ES, a within-campaign early-stop rule that rolls forward the peak checkpoint and sets the next budget to peak_step+3; and GRPO, which changes the RL update using group-relative reward normalization. The answer is regime-dependent. On Qwen-2.5-3B, where naive REINFORCE is fragile, CARE v2 nearly doubles end-of-chain pass@1 from 4.9% to 9.5%, with paired bootstrap 95% CI [+0.4,+8.9] and gains in 4/5 seeds. On Qwen-2.5-7B, CARE reaches parity with naive REINFORCE, 13.8% vs. 11.8%, while ES reaches 22.2% [14.1,28.0]. Out-of-the-box GRPO reaches 20.7% [15.7,25.1], nearly matching REINFORCE+ES. GRPO raises the floor but does not remove the cliff. Its 7B gain mainly comes from better between-campaign carryover, while the within-campaign peak-to-end gap remains about 17 points under both REINFORCE and GRPO. GRPO+ES gives mixed evidence: 2/3 seeds improve, but one final cliff lowers the mean to 17.0% [0.0,28.1]. A Gemma-3-4B pilot shows the same signature, suggesting the phenomenon is not limited to Qwen.

Robots that learn over long deployments must add new skills without losing the shared policy structure that makes earlier skills reusable. We study sequential robot skill learning, where previous trajectories and task losses may be unavailable, and the deployed policy must remain a single shared controller without task-specific heads, routing, or adapters. We identify skill-coupling collapse, a failure mode in which individual skill success remains non-trivial while reliability among related skills deteriorates. We propose Sleeping Robots, a wake-sleep framework that learns each new skill during wake and consolidates the shared policy offline during sleep using compact frozen skill memories: frozen critics with unordered state buffers for reinforcement learning and frozen actor snapshots with unordered observation buffers for imitation learning. During sleep, these memories define differentiable surrogate objectives whose gradients are combined through Nash bargaining, with adaptive anchoring and local excitability for stable consolidation. On Meta-World MT5, Sleeping Robots improves average success by 64 % and pairwise reliability by x 2.0 over the strongest non-oracle baseline, and on SurgicAI it improves average success and backward transfer relative to continual imitation baselines while remaining competitive on pairwise reliability.

Reinforcement Learning (RL) has received increasing attention and adoption in real-world use cases. Most of these systems follow a train-then-fix paradigm, where trained agents do not learn while interacting with the world until performance degrades and retraining becomes necessary. In this position paper, we argue that deploying an agent that is incapable of optimality, but receives an evaluative reward signal, is inherently a continual RL problem. We identify four sources of non-stationarity after deployment that necessitate never-ending learning, and highlight why the best deployed agents never stop adapting. We analyze successful examples of continual RL in the real world, and present the community with the advantages and measures to move away from the current train-then-fix paradigm.

Continual reinforcement learning (RL) aims to produce agents that never stop adapting to new tasks. A key question is how this interacts with the diversity of tasks an agent experiences. Prior work has shown that training on many diverse tasks leads to agents with strong zero-shot and in-context adaptation. However, this work evaluated agents after they'd stopped learning, i.e. with frozen weights. How task diversity affects an agent's ability to continue learning over a sequence of distribution shifts remains unclear. We introduce Banyan, a GPU-accelerated continual RL domain where one can parametrically control three independent axes that define a task: the map layouts an agent must navigate, the objects it must interact with, and the hierarchical structures of sub-goal dependencies. We find that increasing diversity along each axis induces systematic transfer -- that is, agents begin training on a new task distribution near the performance attained on the previous one, even when the shift changes the structure of the optimal policy. While increasing diversity improves systematic transfer, we find that too much diversity inhibits a learner's ability to continue adapting to new task distributions. As diversity increases, learners plateau in the success rate they achieve on new tasks, yet continue improving on old tasks -- even without further exposure to them. We find this phenomenon manifests across continual learning algorithms, memory architectures, architecture sizes, and in Kinetix -- a physics-based control domain. We release Banyan as a domain for running controlled experiments that study continual RL in the many-tasks regime. Code is available at https://github.com/nhshah15/banyan.

Reinforcement learning (RL) with verifiable rewards has achieved strong progress in reasoning-oriented LLMs, but extending it to multi-domain RL remains challenging due to reward unreliability in non-verifiable tasks and capability interference across domains. We propose CARE-RL to combine protocol-aware reward generation with capability-aware optimization for mitigating cross-domain conflicts. For non-verifiable tasks, the Protocol-Aware Generative Reward Model (PA-GRM) constructs prompt-level evaluation protocols and schemas before producing trace-conditioned rewards, enabling task-adaptive yet comparable evaluation of open-ended responses. For multi-domain optimization, Direction-Aware Capability Subspace Projection (DACSP) extracts historical capability directions from previous RL stages and modulates later updates by amplifying aligned components, suppressing conflicting components, and preserving orthogonal updates. Experiments across math, chat, and instruction-following benchmarks show that CARE-RL consistently outperforms standard multi-domain RL baselines, achieving Total Avg scores of 47.9 and 50.7 on Qwen2.5-7B and Qwen3-4B, respectively.

A hallmark of intelligence is the ability to adapt in non-stationary environments, yet deep Reinforcement Learning (RL) agents often struggle in such settings. Prior studies introduce non-stationarity through abrupt shifts in features or dynamics, whereas real-world environments often evolve gradually through continual drift. This distinction has important implications for the "stability-plasticity dilemma" in RL, as abrupt task changes may demand more plasticity than naturalistic settings. To address this, we modify existing 3D Miniworld and MuJoCo environments to incorporate naturalistic, continual non-stationarity, and use them to examine how stability and adaptation affect performance under continuous environmental change. We find that methods favoring stability, such as synaptic consolidation, outperform approaches focused on plasticity, such as parameters resetting. Motivated by this result, and prior evidence that Successor Features (SFs) reduce interference, we investigate whether SFs are better consolidation targets than Q-values. Across both environments, applying neuro-inspired synaptic consolidation to SFs yields superior performance on continually changing settings. Moreover, consolidation is most effective when SFs are stabilized across multiple timescales, which capture complementary aspects of gradual environmental change. Together, these results suggest that stability is more critical in continual learning when changes are gradual, and that multi-timescale consolidation of predictive representations is an effective approach.

Data rehearsal has emerged as a leading approach for mitigating catastrophic forgetting in Continual Reinforcement Learning (CRL). However, existing work remains confined to policy gradient frameworks, regularizing only actors due to the performance degradation incurred by critic regularization. This actor-centric approach overlooks the potential of data rehearsal for value function approximation. Moreover, existing evaluations in CRL rarely consider multi-cyclic environments where task sequences repeat, a critical real-world scenario that exacerbates forgetting and plasticity. We investigate data rehearsal for Deep Q-Networks using Q-value regularization in multi-cyclic settings and propose Qreg+NWLU which introduces two simple modifications: (1) continuous data rehearsal that dynamically collects and updates stored Q-values throughout training, and (2) "No-Wait" regularization that applies immediately rather than after the first task. Together, these modifications yield improvements in learning efficiency, forgetting mitigation, and knowledge transfer over Qreg and conventional CRL methods within value function approximation settings.

Safe reinforcement learning in nonstationary environments requires safety mechanisms that adapt as environmental conditions change. Standard safe reinforcement learning methods often assume fixed constraints or stable environmental conditions, which can become inadequate under distribution shift. We propose LILAC+, a framework for safe continual reinforcement learning under nonstationarity that combines three adaptive safety mechanisms: context-based safety constraints, adaptation-speed constraints, and budget-to-state safety enforcement. Context-based constraints adjust safety requirements using inferred and predicted environmental context. Adaptation-speed constraints tighten safety requirements when the rate of environmental change exceeds the agent's ability to adapt safely. Budget-to-state enforcement converts cumulative safety requirements into local state-level control constraints that can be enforced at decision time. Together, these mechanisms provide a unified approach for proactive and reactive safety adaptation in continual reinforcement learning. We evaluate the framework in simulated driving environments under stationary, seen nonstationary, and unseen nonstationary conditions. The results show that adaptive safety constraints substantially reduce safety violations under distribution shift while maintaining competitive task performance compared with unconstrained and fixed-constraint baselines. These findings suggest that safe continual reinforcement learning requires adaptive constraint mechanisms that respond not only to current state information but also to predicted environmental context, adaptation demand, and remaining safety budget.

In deep reinforcement learning (DRL), an agent is trained from a stream of experience. In a continual learning setting, such agents can suffer from plasticity loss: their ability to learn new skills from new experiences diminishes over training. Recently, Mixture-of-Experts (MoE) networks have been reported to enable scaling laws and facilitate the learning of diverse skills. However, in continual reinforcement learning settings, their performance can degenerate as learning proceeds, indicating a loss of plasticity. To address this, building on Neural Tangent Kernel (NTK) theory, we formalize the plasticity loss in MoE policies as a loss of spectral plasticity. We then derive a tractable proxy for spectral plasticity, one expressible in terms of individual expert feature matrices. Leveraging this proxy, we introduce SPHERE, a practical Parseval penalty tailored for MoE-based policies that alleviates the loss of spectral plasticity. On MetaWorld and HumanoidBench, SPHERE improves average success under continual RL by 133% and 50% over an unregularized MoE baseline, while maintaining higher spectral plasticity throughout training.

In continual reinforcement learning (CRL), good performance requires never-ending learning, acting, and exploration in a big, partially observable world. Most CRL experiments have focused on loss of plasticity -- the inability to keep learning -- in one-off experiments where some unobservable non-stationarity is added to classic fully observable MDPs. Further, these experiments rarely consider the role of partial observability and the importance of CRL agents that use memory or recurrence. One potential reason for this focus on mitigating loss of plasticity without considering partial observability is that many partially-observable CRL environments are prohibitively expensive. In this paper, we introduce Forager, a light-weight partially-observable CRL environment with a constant memory footprint. We provide a set of experiments and sample tasks demonstrating that Forager is challenging for current CRL agents and yet also allows for in-depth study of those agents. We demonstrate that agents exhibit loss of plasticity, proposed mitigations can help, but that most useful is to leverage state construction. We conclude with a variant of Forager that generates an unending stream of new tasks to learn that clearly highlights the limitations of current CRL agents.

Continual offline reinforcement learning (CORL) aims to learn a sequence of tasks from datasets collected over time while preserving performance on previously learned tasks. This setting corresponds to domains where new tasks arise over time, but adapting the model in live environment interactions is expensive, risky, or impossible. However, CORL inherits the dual difficulty of offline reinforcement learning and adapting while preventing catastrophic forgetting. Replay-based continual learning approaches remain a strong baseline but incur memory overhead and suffer from a distribution mismatch between replayed samples and newly learned policies. At the same time, architectural continual learning methods have shown strong potential in supervised learning but remain underexplored in CORL. In this work, we propose TSN-Affinity, a novel CORL method based on TinySubNetworks and Decision Transformer. The method enables task-specific parameterization and controlled knowledge sharing through a RL-aware reuse strategy that routes tasks according to action compatibility and latent similarity. We evaluate the approach on benchmarks based on Atari games and simulations of manipulation tasks with the Franka Emika Panda robotic arm, covering both discrete and continuous control. Results show strong retention from sparse SubNetworks, with routing further improving multi-task performance. Our findings suggest that similarity-guided architectural reuse is a strong and viable alternative to replay-based strategies in a CORL setting. Our code is available at: https://github.com/anonymized-for-submission123/tsn-affinity.

Continual reinforcement learning must balance retention with adaptation, yet many methods still rely on \emph{single-model preservation}, committing to one evolving policy as the main reusable solution across tasks. Even when a previously successful policy is retained, it may no longer provide a reliable starting point for rapid adaptation after interference, reflecting a form of \emph{loss of plasticity} that single-policy preservation cannot address. Inspired by quality-diversity methods, we introduce \textsc{TeLAPA} (Transfer-Enabled Latent-Aligned Policy Archives), a continual RL framework that organizes behaviorally diverse policy neighborhoods into per-task archives and maintains a shared latent space so that archived policies remain comparable and reusable under non-stationary drift. This perspective shifts continual RL from retaining isolated solutions to maintaining \emph{skill-aligned neighborhoods} with competent and behaviorally related policies that support future relearning. In our MiniGrid CL setting, \textsc{TeLAPA} learns more tasks successfully, recovers competence faster on revisited tasks after interference, and retains higher performance across a sequence of tasks. Our analyses show that source-optimal policies are often not transfer-optimal, even within a local competent neighborhood, and that effective reuse depends on retaining and selecting among multiple nearby alternatives rather than collapsing them to one representative. Together, these results reframe continual RL around reusable and competent policy neighborhoods, providing a route beyond single-model preservation toward more plastic lifelong agents.

Ordinary decentralized multi-agent reinforcement learning presents each focal agent with a continual learning problem: peer updates change its induced rewards and dynamics even when the joint Markov game is stationary. We connect the lifetime of success-conditioned reusable structure to peer learning and policy reuse. An invariant core represents maximal abstract patterns shared by a high fraction of successful trajectories; survival refers to a fixed pattern's coverage, not an unchanged maximal frontier. An established sharp conditioning bound limits coverage loss to $\frac{\varepsilon}{p_0}$, where $\varepsilon$ is trajectory-law drift, $p_0$ is reference success mass, and drift is smaller than the initial compatible-success mass. Combining this bound with peer-policy movement certifies $Ω(\frac{1}η)$ survival under bounded peer updates of size $η$ and positive coverage margin. An effective-conflict condition yields a matching $Θ(\frac{1}η)$ first-exit law and holds in an analytic exact-policy-gradient class. With success-mass and performance calibration, structural survival also yields policy-value and finite-library transfer guarantees. An exactly solvable corridor tests the structural predictions. Two registered 64-stream studies in continual control and cue-MNIST show that coverage erosion predicts impending failure and enables near-oracle intervention. An exploratory reanalysis of eight learned-partner Level-Based Foraging pairings suggests the same erosion--failure link under peer learning.

While Large Language Model (LLM) agents excel at general tasks, they inherently struggle with continual adaptation due to the frozen weights after deployment. Conventional reinforcement learning (RL) offers a solution but incurs prohibitive computational costs and the risk of catastrophic forgetting. We introduce Just-In-Time Reinforcement Learning (JitRL), a training-free framework that enables test-time policy optimization without any gradient updates. JitRL maintains a dynamic, non-parametric memory of experiences and retrieves relevant trajectories to estimate action advantages on-the-fly. These estimates are then used to directly modulate the LLM's output logits. We theoretically prove that this additive update rule is the exact closed-form solution to the KL-constrained policy optimization objective. Extensive experiments on WebArena and Jericho demonstrate that JitRL establishes a new state-of-the-art among training-free methods. Crucially, JitRL outperforms the performance of computationally expensive fine-tuning methods (e.g., WebRL) while reducing monetary costs by over 30 times, offering a scalable path for continual learning agents. The code is available at https://github.com/liushiliushi/JitRL.

A continual learning agent builds on previous experiences to develop increasingly complex behaviors by adapting to non-stationary and dynamic environments while preserving previously acquired knowledge. However, scaling these systems presents significant challenges, particularly in balancing the preservation of previous policies with the adaptation of new ones to current environments. This balance, known as the stability-plasticity dilemma, is especially pronounced in complex multi-agent domains such as the train scheduling problem, where environmental and agent behaviors are constantly changing, and the search space is vast. In this work, we propose addressing these challenges in the train scheduling problem using curriculum learning. We design a curriculum with adjacent skills that build on each other to improve generalization performance. Introducing a curriculum with distinct tasks introduces non-stationarity, which we address by proposing a new algorithm: Continual Deep Q-Network (DQN) Expansion (CDE). Our approach dynamically generates and adjusts Q-function subspaces to handle environmental changes and task requirements. CDE mitigates catastrophic forgetting through EWC while ensuring high plasticity using adaptive rational activation functions. Experimental results demonstrate significant improvements in learning efficiency and adaptability compared to RL baselines and other adapted methods for continual learning, highlighting the potential of our method in managing the stability-plasticity dilemma in the adaptive train scheduling setting.