In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.

We introduce an offline multi-agent reinforcement learning (offline MARL) framework that utilizes previously collected data without additional online data collection. Our method reformulates offline MARL as a sequence modeling problem and thus builds on top of the simplicity and scalability of the Transformer architecture. In the fashion of centralized training and decentralized execution, we propose to first train a teacher policy who has the privilege to access every agent's observations, actions, and rewards. After the teacher policy has identified and recombined the "good" behavior in the dataset, we create separate student policies and distill not only the teacher policy's features but also its structural relations among different agents' features to student policies. We show that our framework significantly improves performances on a range of tasks and outperforms state-of-the-art offline MARL baselines. Furthermore, we demonstrate that the proposed method has a better convergence rate, is more sample efficient, and is more robust to various demonstration qualities compared with baselines.

The powerful generalization of Vision-Language-Action (VLA) models is bottlenecked by their heavy reliance on massive, redundant, and unevenly valued datasets, hindering their widespread application. Existing model-centric optimization paths, such as model compression (which often leads to performance degradation) or policy distillation (whose products are model-dependent and lack generality), fail to fundamentally address this data-level challenge. To this end, this paper introduces FT-NCFM, a fundamentally different, data-centric generative data distillation framework. Our framework employs a self-contained Fact-Tracing (FT) engine that combines causal attribution with programmatic contrastive verification to assess the intrinsic value of samples. Guided by these assessments, an adversarial NCFM process synthesizes a model-agnostic, information-dense, and reusable data asset. Experimental results on several mainstream VLA benchmarks show that models trained on just 5% of our distilled coreset achieve a success rate of 85-90% compared with training on the full dataset, while reducing training time by over 80%. Our work demonstrates that intelligent data distillation is a highly promising new path for building efficient, high-performance VLA models.

Abstract-- Grasping objects in cluttered environments remains a fundamental yet challenging problem in robotic manipulation. While prior works have explored learning-based synergies between pushing and grasping for two-fingered grippers, few have leveraged the high degrees of freedom (DoF) in dexterous hands to perform efficient singulation for grasping in cluttered settings. In this work, we introduce DexSinGrasp, a unified policy for dexterous object singulation and grasping. DexSinGrasp enables high-dexterity object singulation to facilitate grasping, significantly improving efficiency and effectiveness in cluttered environments. We incorporate clutter arrangement curriculum learning to enhance success rates and generalization across diverse clutter conditions, while policy distillation enables a deploy-able vision-based grasping strategy. T o evaluate our approach, we introduce a set of cluttered grasping tasks with varying object arrangements and occlusion levels. Experimental results show that our method outperforms baselines in both efficiency and grasping success rate, particularly in dense clutter . Dexterous grasping of target objects in cluttered environments is crucial for various applications, from production lines [1] to assembly processes [2], [3] and beyond.

In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.

Value function initialization (VFI) is an effective way to achieve a jumpstart in reinforcement learning (RL) by leveraging value estimates from prior tasks. While this approach is well established in tabular settings, extending it to deep reinforcement learning (DRL) poses challenges due to the continuous nature of the state-action space, the noisy approximations of neural networks, and the impracticality of storing all past models for reuse. In this work, we address these challenges and introduce DQInit, a method that adapts value function initialization to DRL. DQInit reuses compact tabular Q-values extracted from previously solved tasks as a transferable knowledge base. It employs a knownness-based mechanism to softly integrate these transferred values into underexplored regions and gradually shift toward the agent's learned estimates, avoiding the limitations of fixed time decay. Our approach offers a novel perspective on knowledge transfer in DRL by relying solely on value estimates rather than policies or demonstrations, effectively combining the strengths of jumpstart RL and policy distillation while mitigating their drawbacks. Experiments across multiple continuous control tasks demonstrate that DQInit consistently improves early learning efficiency, stability, and overall performance compared to standard initialization and existing transfer techniques.

Developing controllers that generalize across diverse robot morphologies remains a significant challenge in legged locomotion. Traditional approaches either create specialized controllers for each morphology or compromise performance for generality. This paper introduces a two-stage teacher-student framework that bridges this gap through policy distillation. First, we train specialized teacher policies optimized for individual morphologies, capturing the unique optimal control strategies for each robot design. Then, we distill this specialized expertise into a single Transformer-based student policy capable of controlling robots with varying leg configurations. Our experiments across five distinct legged morphologies demonstrate that our approach preserves morphology-specific optimal behaviors, with the Transformer architecture achieving 94.47% of teacher performance on training morphologies and 72.64% on unseen robot designs. Comparative analysis reveals that Transformer-based architectures consistently outperform MLP baselines by leveraging attention mechanisms to effectively model joint relationships across different kinematic structures. We validate our approach through successful deployment on a physical quadruped robot, demonstrating the practical viability of our morphology-agnostic control framework. This work presents a scalable solution for developing universal legged robot controllers that maintain near-optimal performance while generalizing across diverse morphologies.

-- As robots are increasingly deployed in diverse application domains, generalizable cross-embodiment mobility policies are increasingly essential. While classical mobility stacks have proven effective on specific robot platforms, they pose significant challenges when scaling to new embodiments. Learning-based methods, such as imitation learning (IL) and reinforcement learning (RL), offer alternative solutions but suffer from covariate shift, sparse sampling in large environments, and embodiment-specific constraints. This paper introduces COMPASS, a novel workflow for developing cross-embodiment mobility policies by integrating IL, residual RL, and policy distillation. We begin with IL on a mobile robot, leveraging easily accessible teacher policies to train a foundational model that combines a world model with a mobility policy. Building on this base, we employ residual RL to fine-tune embodiment-specific policies, exploiting pre-trained representations to improve sampling efficiency in handling various physical constraints and sensor modalities. We empirically demonstrate that COMPASS scales effectively across diverse robot platforms while maintaining adaptability to various environment configurations, achieving a generalist policy with a success rate approximately 5X higher than the pre-trained IL policy. The resulting framework offers an efficient, scalable solution for cross-embodiment mobility, enabling robots with different designs to navigate safely and efficiently in complex scenarios.

The advancement in autonomous vehicles has empowered navigation and exploration in unknown environments. Geomagnetic navigation for autonomous vehicles has drawn increasing attention with its independence from GPS or inertial navigation devices. While geomagnetic navigation approaches have been extensively investigated, the generalizability of learned geomagnetic navigation strategies remains unexplored. The performance of a learned strategy can degrade outside of its source domain where the strategy is learned, due to a lack of knowledge about the geomagnetic characteristics in newly entered areas. This paper explores the generalization of learned geomagnetic navigation strategies via deep reinforcement learning (DRL). Particularly, we employ DRL agents to learn multiple teacher models from distributed domains that represent dispersed navigation strategies, and amalgamate the teacher models for generalizability across navigation areas. We design a reward shaping mechanism in training teacher models where we integrate both potential-based and intrinsic-motivated rewards. The designed reward shaping can enhance the exploration efficiency of the DRL agent and improve the representation of the teacher models. Upon the gained teacher models, we employ multi-teacher policy distillation to merge the policies learned by individual teachers, leading to a navigation strategy with generalizability across navigation domains. We conduct numerical simulations, and the results demonstrate an effective transfer of the learned DRL model from a source domain to new navigation areas. Compared to existing evolutionary-based geomagnetic navigation methods, our approach provides superior performance in terms of navigation length, duration, heading deviation, and success rate in cross-domain navigation. Geomagnetic navigation leverages the ubiquitous earth magnetic field signals for the navigation [1], [2], without independence on dedicated devices along the navigation route [3]-[5]. Geomagnetic navigation thus can secure the navigation mission, e.g., in remote areas or underwater environments where there GPS or pre-deployed navigation devices is unavailable [6].

Weiyao Wang 1 and Gregory D. Hager 1 Abstract -- Deploying machine learning algorithms for robot tasks in real-world applications presents a core challenge: overcoming the domain gap between the training and the deployment environment. This is particularly difficult for vi-suomotor policies that utilize high-dimensional images as input, particularly when those images are generated via simulation. A common method to tackle this issue is through domain randomization, which aims to broaden the span of the training distribution to cover the test-time distribution. However, this approach is only effective when the domain randomization encompasses the actual shifts in the test-time distribution. We take a different approach, where we make use of a single demonstration (a prompt) to learn policy that adapts to the testing target environment. Our proposed framework, PromptAdapt, leverages the Transformer architecture's capacity to model sequential data to learn demonstration-conditioned visual policies, allowing for in-context adaptation to a target domain that is distinct from training. Our experiments in both simulation and real-world settings show that PromptAdapt is a strong domain-adapting policy that outperforms baseline methods by a large margin under a range of domain shifts, including variations in lighting, color, texture, and camera pose. Videos and more information can be viewed at project webpage: https://sites.google.com/view/promptadapt.