Achieving expressive and generalizable whole-body motion control is essential for deploying humanoid robots in real-world environments. In this work, we propose UniTracker, a three-stage training framework that enables robust and scalable motion tracking across a wide range of human behaviors. In the first stage, we train a teacher policy with privileged observations to generate high-quality actions. In the second stage, we introduce a Conditional Variational Autoencoder (CVAE) to model a universal student policy that can be deployed directly on real hardware. The CVAE structure allows the policy to learn a global latent representation of motion, enhancing generalization to unseen behaviors and addressing the limitations of standard MLP-based policies under partial observations. Unlike pure MLPs that suffer from drift in global attributes like orientation, our CVAE-student policy incorporates global intent during training by aligning a partial-observation prior to the full-observation encoder. In the third stage, we introduce a fast adaptation module that fine-tunes the universal policy on harder motion sequences that are difficult to track directly. This adaptation can be performed both for single sequences and in batch mode, further showcasing the flexibility and scalability of our approach. We evaluate UniTracker in both simulation and real-world settings using a Unitree G1 humanoid, demonstrating strong performance in motion diversity, tracking accuracy, and deployment robustness.

The development of a generalist agent capable of solving a wide range of sequential decision-making tasks remains a significant challenge. We address this problem in a cross-agent setup where agents share the same observation space but differ in their action spaces. Our approach builds on the universal policy framework, which decouples policy learning into two stages: a diffusion-based planner that generates observation sequences and an inverse dynamics model that assigns actions to these plans. We propose a method for training the planner on a joint dataset composed of trajectories from all agents. This method offers the benefit of positive transfer by pooling data from different agents, while the primary challenge lies in adapting shared plans to each agent's unique constraints. We evaluate our approach on the BabyAI environment, covering tasks of varying complexity, and demonstrate positive transfer across agents. Additionally, we examine the planner's generalisation ability to unseen agents and compare our method to traditional imitation learning approaches. By training on a pooled dataset from multiple agents, our universal policy achieves an improvement of up to $42.20\%$ in task completion accuracy compared to a policy trained on a dataset from a single agent.

This work presents a deep reinforcement learning-based approach to develop a policy for robot-agnostic locomotion control. Our method involves training an agent equipped with memory, implemented as a recurrent policy, on a diverse set of procedurally generated quadruped robots. We demonstrate that the policies trained by our framework transfer seamlessly to both simulated and real-world quadrupeds not encountered during training, maintaining high-quality motion across platforms. Through a series of simulation and hardware experiments, we highlight the critical role of the recurrent unit in enabling generalization, rapid adaptation to changes in the robot's dynamic properties, and sample efficiency.

Learning a universal manipulation policy encompassing doors with diverse categories, geometries and mechanisms, is crucial for future embodied agents to effectively work in complex and broad real-world scenarios. Due to the limited datasets and unrealistic simulation environments, previous works fail to achieve good performance across various doors. In this work, we build a novel door manipulation environment reflecting different realistic door manipulation mechanisms, and further equip this environment with a large-scale door dataset covering 6 door categories with hundreds of door bodies and handles, making up thousands of different door instances. Additionally, to better emulate real-world scenarios, we introduce a mobile robot as the agent and use the partial and occluded point cloud as the observation, which are not considered in previous works while possessing significance for real-world implementations. To learn a universal policy over diverse doors, we propose a novel framework disentangling the whole manipulation process into three stages, and integrating them by training in the reversed order of inference. Extensive experiments validate the effectiveness of our designs and demonstrate our framework's strong performance. Code, data and videos are avaible on https://unidoormanip.github.io/.

We introduce a new programming paradigm called oracle-guided decision programming in which a program specifies a Markov Decision Process (MDP) and the language provides a universal policy. We prototype a new programming language, Dodona, that manifests this paradigm using a primitive 'choose' representing nondeterministic choice. The Dodona interpreter returns either a value or a choicepoint that includes a lossless encoding of all information necessary in principle to make an optimal decision. Meta-interpreters query Dodona's (neural) oracle on these choicepoints to get policy and value estimates, which they can use to perform heuristic search on the underlying MDP. We demonstrate Dodona's potential for zero-shot heuristic guidance by meta-learning over hundreds of synthetic tasks that simulate basic operations over lists, trees, Church datastructures, polynomials, first-order terms and higher-order terms.

Transfer and adaptation to new unknown environmental dynamics is a key challenge for reinforcement learning (RL). An even greater challenge is performing near-optimally in a single attempt at test time, possibly without access to dense rewards, which is not addressed by current methods that require multiple experience rollouts for adaptation. To achieve single episode transfer in a family of environments with related dynamics, we propose a general algorithm that optimizes a probe and an inference model to rapidly estimate underlying latent variables of test dynamics, which are then immediately used as input to a universal control policy. This modular approach enables integration of state-of-the-art algorithms for variational inference or RL. Moreover, our approach does not require access to rewards at test time, allowing it to perform in settings where existing adaptive approaches cannot. In diverse experimental domains with a single episode test constraint, our method significantly outperforms existing adaptive approaches and shows favorable performance against baselines for robust transfer.