Humanoid locomotion faces a critical scalability challenge: traditional reinforcement learning (RL) methods require task-specific rewards and struggle to leverage growing datasets, even as more training terrains are introduced. We propose DreamPolicy, a unified framework that enables a single policy to master diverse terrains and generalize zero-shot to unseen scenarios by systematically integrating offline data and diffusion-driven motion synthesis. At its core, DreamPolicy introduces Humanoid Motion Imagery (HMI) - future state predictions synthesized through an autoregressive terrain-aware diffusion planner curated by aggregating rollouts from specialized policies across various distinct terrains. Unlike human motion datasets requiring laborious retargeting, our data directly captures humanoid kinematics, enabling the diffusion planner to synthesize "dreamed" trajectories that encode terrain-specific physical constraints. These trajectories act as dynamic objectives for our HMI-conditioned policy, bypassing manual reward engineering and enabling cross-terrain generalization. DreamPolicy addresses the scalability limitations of prior methods: while traditional RL fails to exploit growing datasets, our framework scales seamlessly with more offline data. As the dataset expands, the diffusion prior learns richer locomotion skills, which the policy leverages to master new terrains without retraining. Experiments demonstrate that DreamPolicy achieves average 90% success rates in training environments and an average of 20% higher success on unseen terrains than the prevalent method. It also generalizes to perturbed and composite scenarios where prior approaches collapse. By unifying offline data, diffusion-based trajectory synthesis, and policy optimization, DreamPolicy overcomes the "one task, one policy" bottleneck, establishing a paradigm for scalable, data-driven humanoid control.

Abstract-- Unified models capable of solving a wide variety of tasks have gained traction in vision and NLP due to their ability to share regularities and structures across tasks, which improves individual task performance and reduces computational footprint. However, the impact of such models remains limited in embodied learning problems, which present unique challenges due to interactivity, sample inefficiency, and sequential task presentation. In this work, we present PolyTask, a novel method for learning a single unified model that can solve various embodied tasks through a'learn then distill' mechanism. In the'learn' step, PolyTask leverages a few demonstrations for each task to train task-specific policies. Then, in the'distill' step, task-specific policies are distilled into a single policy using a new distillation method called Behavior Distillation. Given a unified policy, individual task behavior can be extracted through conditioning variables. PolyTask is designed to be conceptually simple while being able to leverage well-established algorithms in RL to enable interactivity, a handful of expert demonstrations to allow for sample efficiency, and preventing interactive access to tasks during distillation to enable lifelong learning. Distillation allows our unified I. Once trained, the unified policy can solve tasks tasks [1, 2, 3, 4]. In contrast to task-specific models, by conditioning on task identifiers such as goal image, text unified models are hypothesized to benefit from sharing data, description, or one-hot labels.

An attached arm can significantly increase the applicability of legged robots to several mobile manipulation tasks that are not possible for the wheeled or tracked counterparts. The standard hierarchical control pipeline for such legged manipulators is to decouple the controller into that of manipulation and locomotion. However, this is ineffective. It requires immense engineering to support coordination between the arm and legs, and error can propagate across modules causing non-smooth unnatural motions. It is also biological implausible given evidence for strong motor synergies across limbs. In this work, we propose to learn a unified policy for whole-body control of a legged manipulator using reinforcement learning. We propose Regularized Online Adaptation to bridge the Sim2Real gap for high-DoF control, and Advantage Mixing exploiting the causal dependency in the action space to overcome local minima during training the whole-body system. We also present a simple design for a low-cost legged manipulator, and find that our unified policy can demonstrate dynamic and agile behaviors across several task setups. Videos are at https://maniploco.github.io