Zero-Shot Policy Transfer

Successor-feature-based zero-shot reinforcement learning methods pre-train a task encoder and successor-feature predictor on offline, reward-free data, then recover a policy at test time by solving a closed-form linear regression over a small labeled dataset. We identify an underexplored limitation of this approach: when the task encoder produces dense representations, the inferred reward signal becomes diffuse across the state space, reducing the contrast between goal-relevant states and the background and providing a weaker signal for the policy to follow. We address this limitation by introducing an l1 sparsity penalty that concentrates reward-predictive structure into a smaller effective subspace of the task encoder's latent space. The penalty is applied only during training and requires no architectural change or modification to the closed-form inference procedure. We instantiate this approach as Sparse-TD-JEPA, a single-line of code addition based on Temporal-Difference Joint-Embedding Predictive Architecture (TD-JEPA) that significantly improved its performance in zero-shot reinforcement learning. We evaluate Sparse-TD-JEPA on 65 tasks across the DeepMind Control (DMC) Suite and OGBench, covering locomotion, navigation, trajectory stitching, puzzle solving, and robotic manipulation. Sparse-TD-JEPA improves over TD-JEPA in all 13 domain-level comparisons. On DMC, it increases the average return from 661 +- 6 to 714 +- 3. On OGBench, it increases the average success rate from 38.0 +- 0.8 to 48.4 +- 2.3, with improvements of 13.4-18.0 percentage points on the evaluated antmaze navigation and stitching domains. Overall, Sparse-TD-JEPA achieves the highest mean performance among established zero-shot baselines in nine of the 13 domains. Analysis further shows that sparsity produces a sharper and more task-informative inferred reward landscape.

This paper explores a reward-based policy to achieve zero-shot transfer between source and target environments with completely different observation spaces. While humans can demonstrate impressive adaptation capabilities, deep neural network policies often struggle to adapt to a new environment and require a considerable amount of samples for successful transfer. Instead, we propose a novel reward-based policy only conditioned on rewards and actions, enabling zero-shot adaptation to new environments with completely different observations. We discuss the challenges and feasibility of a reward-based policy and then propose a practical algorithm for training. We demonstrate that a reward policy can be trained within three different environments, Pointmass, Cartpole, and 2D Car Racing, and transferred to completely different observations, such as different color palettes or 3D rendering, or Stretch robot navigation in Habitat-Sim, in a zero-shot manner. We also demonstrate that a reward-based policy can further guide the training of an observation-based policy in the target environment.

Zero-shot Transfer in Reinforcement Learning (RL) aims to train an agent that can generate optimal policies for any reward function, without additional learning at transfer time, while training only on reward-free trajectories. For their generality over tasks, such models are sometimes called ``Behavioral Foundation Models'' (BFMs). While they have shown strong performances and improvements in recent years, the current framework and algorithms still assume that, during the transfer phase, the agent is informed offline about the reward (the task to solve) through a dataset of state-reward pairs, which it uses to pick the best policy to deploy. However, in practice if the reward is a black-box (e.g. direct user feedback), it is not possible to generate such a dataset: it is necessary to observe the reward through interactions with the environment. In other words, the current framework of offline transfer is not aligned with the traditional RL setting of online learning through trial-and-error, which requires exploration in order to find rewards. This paper proposes to tackle this new online transfer in zero-shot RL, with the key insight that the BFM itself can be used to generate exploration policies. We show that it is possible to frame this online learning problem in terms of a bandit-like exploration-exploitation problem. More precisely, at each step the bandit algorithm recommends a policy, the BFM executes it in the environment, which yields a reward and a new state; we repeat the process until we converge to the optimal policy. In the popular context of linear reward approximation, we derive a formulation inspired by Upper Confidence Bound and show that exploration can be achieved through the minimization of the eigenvalues of an uncertainty matrix. We evaluate qualitatively and quantitatively our framework on a simple environment to validate the concept of our method.

As autonomous aircraft are introduced at scale and traffic density increases, centralized management becomes insufficient to coordinate the large numbers of crewed and uncrewed aircraft. Dedicated Advanced Air Mobility (AAM) corridors have therefore been proposed for organizing high-density autonomous traffic flows. The desire to scalably provide autonomous aircraft flexibility in trajectory planning motivates the development of decentralized approaches to traffic management in AAM corridors. In this work, we extend a multi-agent reinforcement learning (MARL) approach to address the challenge of decentralized traffic flow management in air corridor networks. We test policies trained in a single-corridor setting on increasingly complex multi-corridor networks with combinations of merges and splits in a zero-shot manner. Experimental results demonstrate that learned behaviors transfer well to scenarios with varying traffic density, network geometry, and heterogeneous vehicle performance, without needing centralized coordination or model retraining. We evaluate system-level performance in terms of conformance to corridor boundaries, completion rates, average speeds, distance traveled, and maintenance of inter-aircraft separation. We find that although our policies require only locally coordinated entry, traversal, and exit behaviors, they collectively produce desirable traffic flows through the corridor network.

Real-world dynamics shifts pose a critical challenge for reinforcement learning in robotics, as policies tightly coupled to nominal environments often fail catastrophically when physical conditions change. Most existing methods rely on encoding explicitly identified physical parameters into a latent context, a parameter-centric paradigm that depends on pre-specified axes of variation and becomes brittle under unmodeled or compound dynamics changes. We revisit dynamics adaptation from an outcome-centric perspective: rather than telling policies what the dynamics are, we enable them to learn how dynamics affect interaction outcomes. Theoretically, this is grounded in a monotonic relationship between target-domain regret and the Lipschitz constant of a trajectory dynamics encoder. Practically, this constant can be upper-bounded through contrastive learning, yielding a smooth, task-relevant latent topology without privileged dynamics information. On MuJoCo benchmarks, our method consistently outperforms parameter-centric baselines under severe dynamics shifts, including unmodeled and time-varying parameters, while also improving in-distribution stability and latent interpretability. Overall, these results validate that controlling latent geometry is a principled mechanism for robust adaptation.

We describe an adaptive market-making architecture that preserves the analytical structure of the Avellaneda--Stoikov framework while introducing a successor measure-style adaptation mechanism. In our paper we keep Avellaneda--Stoikov fast Hamilton--Jacobi--Bellman structure and make it adaptive to changing market regimes and trading objectives. The central idea is to separate market dynamics from the trading objective. The market state determines a low-dimensional set of Avellaneda--Stoikov parameters, while recent realized rewards determine a low-dimensional objective vector. The HJB forward map then converts this objective into optimal bid and ask quotes through a scalarization of future reward features.

We formally establish the equivalence between Observation Delay (OD) and Action Delay (AD) in cooperative partially observable multi-agent systems using observation-action histories. We show that both systems generate identical admissible joint-policy sets, and their induced state-action-observation trajectories are identical in distribution, leading to identical optimal solutions in Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs). This formally generalizes existing infinite-horizon single-agent results to any-horizon partially observable cooperative multi-agent problems with decentralized policy execution, and allows any mixed-delay configuration to be reduced to a pure OD system. We further prove that in Transition-Independent MDPs (TI-MDPs), the observation-action history reduces to a tractable minimal local augmented state. However, we show through numerical experiments that although the optimal solution spaces are structurally isomorphic, the practical learning dynamics are fundamentally different. First, using the minimal local augmented state, the equivalence no longer holds when transitions are not independent. Second, operational constraints and causal credit-assignment errors in Temporal Difference (TD) algorithms induce different learning behaviors across regimes. Finally, leveraging this structural equivalence to bypass these learning challenges, we demonstrate successful multi-agent zero-shot policy transfer from OD to AD, paving the way for unified, efficient solution methods in complex delayed systems.

Offline zero-shot reinforcement learning (RL) aims to learn agents that optimize unseen reward functions without additional environment interaction. The standard approach to this problem trains task-conditioned policies by sampling task vectors that define linear reward functions over learned state representations. In most existing algorithms, these task vectors are randomly sampled, implicitly assuming this adequately captures the structure of the task space. We argue that doing so leads to suboptimal zero-shot generalization. To address this limitation, we propose extracting task vectors directly from the offline dataset and using them to define the task distribution used for policy training. We introduce a simple and general reward function extraction procedure that integrates into existing offline zero-shot RL algorithms. Across multiple benchmark environments and baselines, our approach improves zero-shot performance by an average of 20%, highlighting the importance of principled task sampling in offline zero-shot RL.

Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic semantic operator at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent's original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available here.

Generalizing learned locomotion policies across quadrupedal robots with different morphologies remains a challenge. Policies trained on a single robot often fail when deployed on embodiments with different mass distributions, kinematics, joint limits, or actuation constraints, forcing per-robot retraining. Prior works have approached this primarily either by scaling via training across real or generated embodiments or using large architectures producing transferable policies which are both storage and compute heavy. We argue that both of these approaches circumvent a key failure mode in actor-critic learning: a shared value function tends to average incompatible value targets across embodiments, yielding miscalibrated advantages and fixing that helps transfer and reduce the compute cost. We present MorFiC, a reinforcement learning approach for zero-shot cross-morphology locomotion which fixes this issue using multiplicative critic conditioning on a morphology latent. Trained with a single source robot with morphology randomization in simulation, MorFiC achieves zero-shot transfer to total seven robots and reaching forward velocity competitive or surpassing scaling-based and morphology-conditioned PPO baselines for examples 1.98 m/s on AlienGo, where as additive PPO baselines remain below 0.65 m/s. MorFiC achieves this using single training robot within 2 hours of training time compared to tens or hundreds of hours required for scaling baselines training on multiple embodiments. We diagnose critic calibration via three metric: Explained variance, advantage sign-flip rate and policy gradient cosine, confirming our critic conditioning as determinant for policy update direction. Finally, we demonstrate zero-shot deployment on Unitree Go1 and Go2 robots without fine-tuning.

We present PRISM (Policy Reuse via Interpretable Strategy Mapping), a framework that grounds reinforcement learning agents' decisions in discrete, causally validated concepts and uses those concepts as a zero-shot transfer interface between agents trained with different algorithms. PRISM clusters each agent's encoder features into $K$ concepts via K-means. Causal intervention establishes that these concepts directly drive - not merely correlate with - agent behavior: overriding concept assignments changes the selected action in 69.4% of interventions ($p = 8.6 \times 10^{-86}$, 2500 interventions). Concept importance and usage frequency are dissociated: the most-used concept (C47, 33.0% frequency) causes only a 9.4% win-rate drop when ablated, while ablating C16 (15.4% frequency) collapses win rate from 100% to 51.8%. Because concepts causally encode strategy, aligning them via optimal bipartite matching transfers strategic knowledge zero-shot. On Go~7$\times$7 with three independently trained agents, concept transfer achieves 69.5%$\pm$3.2% and 76.4%$\pm$3.4% win rate against a standard engine across the two successful transfer pairs (10 seeds), compared to 3.5% for a random agent and 9.2% without alignment. Transfer succeeds when the source policy is strong; geometric alignment quality predicts nothing ($R^2 \approx 0$). The framework is scoped to domains where strategic state is naturally discrete: the identical pipeline on Atari Breakout yields bottleneck policies at random-agent performance, confirming that the Go results reflect a structural property of the domain.

Current safety mechanisms for Large Language Models (LLMs) rely heavily on static, fine-tuned classifiers that suffer from adaptation rigidity, the inability to enforce new governance rules without expensive retraining. To address this, we introduce CourtGuard, a retrieval-augmented multi-agent framework that reimagines safety evaluation as Evidentiary Debate. By orchestrating an adversarial debate grounded in external policy documents, CourtGuard achieves state-of-the-art performance across 7 safety benchmarks, outperforming dedicated policy-following baselines without fine-tuning. Beyond standard metrics, we highlight two critical capabilities: (1) Zero-Shot Adaptability, where our framework successfully generalized to an out-of-domain Wikipedia Vandalism task (achieving 90% accuracy) by swapping the reference policy; and (2) Automated Data Curation and Auditing, where we leveraged CourtGuard to curate and audit nine novel datasets of sophisticated adversarial attacks. Our results demonstrate that decoupling safety logic from model weights offers a robust, interpretable, and adaptable path for meeting current and future regulatory requirements in AI governance.

Humans excel at analogical reasoning - applying knowledge from one task to a related one with minimal relearning. In contrast, reinforcement learning (RL) agents typically require extensive retraining even when new tasks share structural similarities with previously learned ones. In this work, we propose MAGIK, a novel framework that enables RL agents to transfer knowledge to analogous tasks without interacting with the target environment. Our approach leverages an imagination mechanism to map entities in the target task to their analogues in the source domain, allowing the agent to reuse its original policy. Experiments on custom MiniGrid and MuJoCo tasks show that MAGIK achieves effective zero-shot transfer using only a small number of human-labelled examples. We compare our approach to related baselines and highlight how it offers a novel and effective mechanism for knowledge transfer via imagination-based analogy mapping.

In the zero-shot policy transfer (ZSPT) setting for contextual Markov decision processes (CMDP), agents train on a fixed, finite set of contexts and must generalize to new ones. Recent work has demonstrated that training on additional states, even if they are irrelevant for solving the current context, can improve generalization to unseen contexts. In this paper, we demonstrate that training on these states can indeed improve generalization, but can come at a cost of reducing the accuracy of the learned value function, which should hurt generalization. We hypothesize and demonstrate that increasing the agent's coverage by training on these additional states while also increasing the accuracy improves generalization even further. Inspired by this, we propose a simple approach Explore-Go that leverages existing pure exploration strategies in a new way: by introducing a pure exploration phase at the start of each training episode. Unlike previous approaches that apply exploration strategies for the purpose of improving generalization, our approach can be combined with both on- and off-policy algorithms. We demonstrate the effectiveness of Explore-Go when combined with several popular algorithms and show an increase in test-time performance across several generalization benchmarks, even partially observable ones. With this, we hope to provide practitioners with a simple modification that can significantly improve the generalization of their agents.