Policy Distillation

Skill memories can improve agentic reinforcement learning by reusing past experience as textual guidance, but retrieved skills are not oracular: they may help in one state while misleading the same policy in another. This makes the common privileged-teacher assumption fragile, namely that a skill-conditioned prompt can be treated as a fixed teacher for the no-skill prompt. We introduce UCOB, a framework for learning to utilize and evolve agentic skills via credit-aware on-policy bidirectional self-distillation. UCOB treats skill-conditioned and no-skill prompts as two on-policy context views of the same model, compares their return-to-go within the same task and anchor state, and uses the higher-return view as the local teacher. This local credit signal internalizes useful skill-conditioned behavior, corrects misleading skill usage, and guides task/state skill memory updates, utility-aware retrieval, and reflection self-training. Experiments on agentic tasks, including ALFWorld, WebShop, and Search-QA, show that UCOB outperforms skill-free RL, skill-memory baselines, and self-distillation methods across model scales, with up to 23.5 and 18.0 point gains over SOTA baselines on ALFWorld and WebShop. Ablations and analyses further validate its core mechanisms and efficiency.

Humanoid robots have achieved strong locomotion capabilities, but reliable navigation on versatile terrains remains challenging because obstacle avoidance must be coordinated with dynamically feasible motion. In this work, we present GuideWalk, a unified end-to-end framework that integrates traversability-aware navigation guidance with terrain-adaptive locomotion teacher for humanoid navigation. Specifically, we introduce a navigation module that provides explicit velocity guidance, decoupling obstacle avoidance from terrain conditions to enable robust planning across diverse environments. We propose a composite teacher distillation scheme, where goal-directed commands and dynamically consistent actions are aggregated and distilled into a single policy. To further improve robustness, the distilled policy is refined with reinforcement learning and an auxiliary behavior cloning objective, which promotes exploration while preserving desirable teacher behaviors. Experiments demonstrate that GuideWalk achieves stable and effective navigation while maintaining stable humanoid locomotion.

Constructing efficient and reliable policies to assist humans is indispensable for human-AI collaboration. Existing methods mainly follow two lines of work. Most prior work relies on multi-agent reinforcement learning (MARL) to learn black-box policies, which limits interpretability and raises safety concerns. Recent methods query large language models (LLMs) at each decision step, causing slow responses and high inference costs. We propose Collaboration Policy Tree (Co-pi-tree), a closed-loop method that learns an executable policy tree consisting of a partner-behavior prediction tree and an agent-action selection tree. Co-pi-tree constructs a policy by distilling LLM reasoning into policy tree code. It then evaluates the policy through partner interaction, obtains feedback, and uses natural language to summarize the interaction feedback to improve problematic branches. Experiments in Overcooked-AI show that Co-pi-tree improves average reward by 35.4% over the baseline average, while reducing the number of LLM queries by 77.7% and test-time latency by 97.1%. Project page: https://beiwenzhang.github.io/Co-pi-tree/

Humanoid loco-manipulation benefits from one controller to coordinate locomotion, arm movements, and fall recovery. These behaviors are difficult to learn together from scratch because they require different skills and objectives. We present HANDOFF, a training architecture that distills motion-tracking, locomotion, and recovery teachers into one 29-DoF policy, using context-conditioned KL losses over separate action slices and a soft mixture of experts. Commanded velocity continuously blends body supervision, while a binary recovery flag assigns the full action to the recovery teacher. The resulting controller takes a compact 10-D task-space command rather than a dense kinematic stream and does not switch policies at runtime. On the Unitree G1, HANDOFF matches state-of-the-art velocity tracking and obtains the largest robust manipulation workspace in a matched adapted-interface benchmark. The same controller executes natural-language-driven, multi-stage loco-manipulation tasks in simulation and hardware experiments without further data collection or fine-tuning.

Deep reinforcement learning (RL) offers a promising route to real-time power grid operation, yet large neural policies are costly to evaluate, hard to deploy on constrained hardware, and opaque to operators. We ask whether a Proximal Policy Optimization (PPO) agent for grid topology control can be compressed into compact tree-based surrogates without losing operational performance. A PPO teacher is trained on Grid2Op's standard 14-bus environment with a stability-oriented reward, using stress-focused data collection on critical, high-loading states. The policy is then distilled into a decision tree and a random forest. Across held-out validation episodes, both surrogates exceed the teacher in mean reward and survival length at a fraction of the inference cost. The decision tree shows high exact-action agreement with the PPO argmax and near-complete agreement within its top-ranked actions, while remaining small enough to be inspected directly. Feature-importance analysis reveals a representational shift: the PPO policy relies mainly on line-loading signals, while the distilled tree is driven primarily by bus-topology variables. These results suggest that stress-focused distillation can convert a black-box neural controller into a lightweight, auditable rule-like surrogate suited for real-time deployment, while also surfacing risks tied to deterministic actions and topology-specific generalization.

Reinforcement learning (RL) is a natural fit for agentic knowledge base question answering (KBQA), where a model must issue executable actions, observe knowledge-base feedback, and eventually return an answer. However, current RL-based KBQA systems mainly optimize sparse rewards from the final answer, leaving intermediate action errors weakly supervised. This is especially limiting for logical-form annotated KBQA benchmarks: gold logical forms can be converted into executable action sequences, but existing pipelines use them mainly for warm-start data construction rather than for on-policy RL updates. We propose GAPD, a training-time Gold-Action Policy Distillation framework that adds dense token-level guidance to outcome-based RL. To align gold actions with on-policy student rollouts, GAPD uses MID-ANCHOR MATCHING: it treats the intermediate entities reached during student exploration and gold execution as state anchors, and matches student states to gold states through these explored entity sets. The current policy conditioned on this aligned gold action serves as a stop-gradient teacher, whose token distribution is distilled back to the ordinary student policy over generated action-token spans. GAPD consistently surpasses the current state of the art on WebQSP, GrailQA, and GraphQ.

Despite many successful attempts at explaining Deep Reinforcement Learning policies using distillation, it remains difficult to balance the performance-interpretability trade-off and select a fitting surrogate model. In addition to this, traditional distillation only minimizes the distance between the behavior of the original and the surrogate policy while other RL-specific components such as action value are disregarded. To solve this, we introduce a new model-agnostic method called Critic-Driven Voronoi State Partitioning, which partitions a black box control policy into regions where a simple class of model can be optimized using gradient descent. By exploiting the critic value network of the original policy, we iteratively introduce new subpolicies in regions with insufficient value, standing in for a measure of policy complexity. The partitioning, a Voronoi quantizer, uses nearest neighbor lookups to assign a linear function to each point in the state space resulting in a cell-like diagram. We validate our approach on several well known benchmarks and proof that this distillation approaches the original policy using a reasonable sized set of linear functions.

Reasoning-centric large language models (LLMs) achieve strong performance by generating intermediate reasoning trajectories, but often incur excessive token usage and high inference-time decoding cost. We observe that, when solving the same problems, larger reasoning models can often produce more concise traces, whereas smaller reasoning models tend to generate longer and more redundant trajectories. This is especially problematic in real-world deployment, where memory, latency, and serving-cost constraints often favor smaller models. Our observations suggest that reasoning compression can be transferred from large models to small ones rather than enforced through explicit length constraints. Based on this insight, we propose Mixed-Policy Distillation (MPD), a reasoning compression framework that transfers concise reasoning behavior from a larger-sized teacher to a smaller student by distilling teacher-compressed student trajectories. Unlike on-policy distillation, which aligns the student with teacher distributions over verbose student trajectories, or off-policy distillation, which relies on teacher-generated trajectories and may suffer from distribution mismatch, MPD combines the strengths of both. Given a student-sampled trajectory, the teacher rewrites it into a more concise reasoning trace, and the student is trained via KL-based alignment on the compressed trajectory. This preserves student-policy exploration while injecting teacher-guided compression. Experiments on Qwen3-1.7B show that MPD reduces token usage by up to 27.1% while improving performance across multiple reasoning benchmarks, demonstrating an effective approach to efficient small-model reasoning.

Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs), whose effectiveness depends on intertwined choices of divergence direction, optimization strategy, and data regime. We break down the design of existing KD methods and present a unified view that establishes connections between them, reformulating KD as a reweighted log-likelihood objective at the token level. We further propose Hybrid Policy Distillation (HPD), which integrates the complementary advantages of forward and reverse KL to balance mode coverage and mode-seeking, and combines off-policy data with lightweight, approximate on-policy sampling. We validate HPD on long-generation math reasoning as well as short-generation dialogue and code tasks, demonstrating improved optimization stability, computational efficiency, and final performance across diverse model families and scales. The code related to this work is available at https://github.com/zwhong714/Hybrid-Policy-Distillation.

Real world deployment of multi agent reinforcement learning MARL systems is fundamentally constrained by limited compute memory and inference time. While expert policies achieve high performance they rely on costly decision cycles and large scale models that are impractical for edge devices or embedded platforms. Knowledge distillation KD offers a promising path toward resource aware execution but existing KD methods in MARL focus narrowly on action imitation often neglecting coordination structure and assuming uniform agent capabilities. We propose resource aware Knowledge Distillation for Multi Agent Reinforcement Learning KD MARL a two stage framework that transfers coordinated behavior from a centralized expert to lightweight decentralized student agents. The student policies are trained without a critic relying instead on distilled advantage signals and structured policy supervision to preserve coordination under heterogeneous and limited observations. Our approach transfers both action level behavior and structural coordination patterns from expert policies while supporting heterogeneous student architectures allowing each agent model capacity to match its observation complexity which is crucial for efficient execution under partial or limited observability and limited onboard resources. Extensive experiments on SMAC and MPE benchmarks demonstrate that KD MARL achieves high performance retention while substantially reducing computational cost. Across standard multi agent benchmarks KD MARL retains over 90 percent of expert performance while reducing computational cost by up to 28.6 times FLOPs. The proposed approach achieves expert level coordination and preserves it through structured distillation enabling practical MARL deployment across resource constrained onboard platforms.

Transformer-based reinforcement learning has emerged as a strong candidate for sequential control in residential energy management. In particular, the Decision Transformer can learn effective battery dispatch policies from historical data, thereby increasing photovoltaic self-consumption and reducing electricity costs. However, transformer models are typically too computationally demanding for deployment on resource-constrained residential controllers, where memory and latency constraints are critical. This paper investigates knowledge distillation to transfer the decision-making behaviour of high-capacity Decision Transformer policies to compact models that are more suitable for embedded deployment. Using the Ausgrid dataset, we train teacher models in an offline sequence-based Decision Transformer framework on heterogeneous multi-building data. We then distil smaller student models by matching the teachers' actions, thereby preserving control quality while reducing model size. Across a broad set of teacher-student configurations, distillation largely preserves control performance and even yields small improvements of up to 1%, while reducing the parameter count by up to 96%, the inference memory by up to 90%, and the inference time by up to 63%. Beyond these compression effects, comparable cost improvements are also observed when distilling into a student model of identical architectural capacity. Overall, our results show that knowledge distillation makes Decision Transformer control more applicable for residential energy management on resource-limited hardware.

Flow-matching policies hold great promise for reinforcement learning (RL) by capturing complex, multi-modal action distributions. However, their practical application is often hindered by prohibitive inference latency and ineffective online exploration. Although recent works have employed one-step distillation for fast inference, the structure of the initial noise distribution remains an overlooked factor that presents significant untapped potential. This overlooked factor, along with the challenge of controlling policy stochasticity, constitutes two critical areas for advancing distilled flow-matching policies. To overcome these limitations, we propose GoldenStart (GSFlow), a policy distillation method with Q-guided priors and explicit entropy control. Instead of initializing generation from uninformed noise, we introduce a Q-guided prior modeled by a conditional VAE. This state-conditioned prior repositions the starting points of the one-step generation process into high-Q regions, effectively providing a "golden start" that shortcuts the policy to promising actions. Furthermore, for effective online exploration, we enable our distilled actor to output a stochastic distribution instead of a deterministic point. This is governed by entropy regularization, allowing the policy to shift from pure exploitation to principled exploration. Our integrated framework demonstrates that by designing the generative startpoint and explicitly controlling policy entropy, it is possible to achieve efficient and exploratory policies, bridging the generative models and the practical actor-critic methods. We conduct extensive experiments on offline and online continuous control benchmarks, where our method significantly outperforms prior state-of-the-art approaches. Code will be available at https://github.com/ZhHe11/GSFlow-RL.

Distilling humanoid locomotion control from offline datasets into deployable policies remains a challenge, as existing methods rely on privileged full-body states that require complex and often unreliable state estimation. We present Sensor-Conditioned Diffusion Policies (SCDP) that enables humanoid locomotion using only onboard sensors, eliminating the need for explicit state estimation. SCDP decouples sensing from supervision through mixed-observation training: diffusion model conditions on sensor histories while being supervised to predict privileged future state-action trajectories, enforcing the model to infer the motion dynamics under partial observability. We further develop restricted denoising, context distribution alignment, and context-aware attention masking to encourage implicit state estimation within the model and to prevent train-deploy mismatch. We validate SCDP on velocity-commanded locomotion and motion reference tracking tasks. In simulation, SCDP achieves near-perfect success on velocity control (99-100%) and 93% tracking success in AMASS test set, performing comparable to privileged baselines while using only onboard sensors. Finally, we deploy the trained policy on a real G1 humanoid at 50 Hz, demonstrating robust real robot locomotion without external sensing or state estimation.

Generative policies based on diffusion and flow matching achieve strong performance in robotic manipulation by modeling multi-modal human demonstrations. However, their reliance on iterative Ordinary Differential Equation (ODE) integration introduces substantial latency, limiting high-frequency closed-loop control. Recent single-step acceleration methods alleviate this overhead but often exhibit distributional collapse, producing averaged trajectories that fail to execute coherent manipulation strategies. We propose a framework that distills a Conditional Flow Matching (CFM) expert into a fast single-step student via Implicit Maximum Likelihood Estimation (IMLE). A bi-directional Chamfer distance provides a set-level objective that promotes both mode coverage and fidelity, enabling preservation of the teacher multi-modal action distribution in a single forward pass. A unified perception encoder further integrates multi-view RGB, depth, point clouds, and proprioception into a geometry-aware representation. The resulting high-frequency control supports real-time receding-horizon re-planning and improved robustness under dynamic disturbances.

Decision transformer based sequential policies have emerged as a powerful paradigm in offline reinforcement learning (RL), yet their efficacy remains constrained by the quality of static datasets and inherent architectural limitations. Specifically, these models often struggle to effectively integrate suboptimal experiences and fail to explicitly plan for an optimal policy. To bridge this gap, we propose Imaginary Planning Distillation (IPD), a novel framework that seamlessly incorporates offline planning into data generation, supervised training, and online inference. Our framework first learns a world model equipped with uncertainty measures and a quasi-optimal value function from the offline data. These components are utilized to identify suboptimal trajectories and augment them with reliable, imagined optimal rollouts generated via Model Predictive Control (MPC). A Transformer-based sequential policy is then trained on this enriched dataset, complemented by a value-guided objective that promotes the distillation of the optimal policy. By replacing the conventional, manually-tuned return-to-go with the learned quasi-optimal value function, IPD improves both decision-making stability and performance during inference. Empirical evaluations on the D4RL benchmark demonstrate that IPD significantly outperforms several state-of-the-art value-based and transformer-based offline RL methods across diverse tasks.

While recent advances have demonstrated strong performance in individual humanoid skills such as upright locomotion, fall recovery and whole-body coordination, learning a single policy that masters all these skills remains challenging due to the diverse dynamics and conflicting control objectives involved. To address this, we introduce X-Loco, a framework for training a vision-based generalist humanoid locomotion policy. X-Loco trains multiple oracle specialist policies and adopts a synergetic policy distillation with a case-adaptive specialist selection mechanism, which dynamically leverages multiple specialist policies to guide a vision-based student policy. This design enables the student to acquire a broad spectrum of locomotion skills, ranging from fall recovery to terrain traversal and whole-body coordination skills. To the best of our knowledge, X-Loco is the first framework to demonstrate vision-based humanoid locomotion that jointly integrates upright locomotion, whole-body coordination and fall recovery, while operating solely under velocity commands without relying on reference motions. Experimental results show that X-Loco achieves superior performance, demonstrated by tasks such as fall recovery and terrain traversal. Ablation studies further highlight that our framework effectively leverages specialist expertise and enhances learning efficiency.

Humans rearrange objects in cluttered environments using egocentric perception, actively moving to keep task-relevant spatial cues in view. Mobile robots have not matched this: rearrangement is usually built on a global pose estimate or a map, which is exactly what a robot carrying one camera lacks, while pushing keeps changing the scene it would have to be built from. We present EgoPush, which pushes objects into anchor-relative formations from onboard RGB-D alone, with no global localization, external tracking, or map at deployment, and transfers zero-shot to a TurtleBot in controlled and visually cluttered scenes. What makes this learnable turns out to be a property of the teacher rather than of the student: three privileged teachers trained with identical rewards, architecture, and hyperparameters all exceed $98\%$ success, yet their distilled egocentric students reach $0\%$, $54.8\%$, and $87.3\%$, the only variable being the teacher's observation function. EgoPush therefore trains the teacher under egocentric observability constraints, restricting it to visibility-limited cues and revealing target references only when the anchor is centrally visible, so that its supervision is recoverable by a depth-based student distilled online. Making the teacher trainable in the first place needs two further pieces: a role-grouped object-centric interface shared by teacher and student, and stage-wise temporally decayed rewards for long-horizon credit assignment. Videos, the playable task, and code are available at https://ai4ce.github.io/EgoPush/.

Partial observability is a notorious challenge in reinforcement learning (RL), due to the need to learn complex, history-dependent policies. Recent empirical successes have used privileged expert distillation--which leverages availability of latent state information during training (e.g., from a simulator) to learn and imitate the optimal latent, Markovian policy--to disentangle the task of "learning to see" from "learning to act". While expert distillation is more computationally efficient than RL without latent state information, it also has well-documented failure modes. In this paper--through a simple but instructive theoretical model called the perturbed Block MDP, and controlled experiments on challenging simulated locomotion tasks--we investigate the algorithmic trade-off between privileged expert distillation and standard RL without privileged information. Our main findings are: (1) The trade-off empirically hinges on the stochasticity of the latent dynamics, as theoretically predicted by contrasting approximate decodability with belief contraction in the perturbed Block MDP; and (2) The optimal latent policy is not always the best latent policy to distill. Our results suggest new guidelines for effectively exploiting privileged information, potentially advancing the efficiency of policy learning across many practical partially observable domains.