Coordinated Humanoid Manipulation with Choice Policies

TL;DR

The paper tackles robust, coordinated whole-body manipulation for humanoids in human-centric environments by coupling a modular teleoperation interface with a fast, multimodal imitation learner called Choice Policy. The policy outputs multiple action proposals and learns to score them, enabling single-pass inference that preserves multimodality while maintaining real-time control. Empirical results on dishwasher loading and whiteboard wiping show Choice Policy outperforming diffusion-based policies and standard behavior cloning, with hand-eye coordination identified as essential for long-horizon tasks. Together, these components illustrate a scalable path for collecting high-quality demonstrations and learning coordinated loco-manipulation policies in unstructured settings.

Abstract

Humanoid robots hold great promise for operating in human-centric environments, yet achieving robust whole-body coordination across the head, hands, and legs remains a major challenge. We present a system that combines a modular teleoperation interface with a scalable learning framework to address this problem. Our teleoperation design decomposes humanoid control into intuitive submodules, which include hand-eye coordination, grasp primitives, arm end-effector tracking, and locomotion. This modularity allows us to collect high-quality demonstrations efficiently. Building on this, we introduce Choice Policy, an imitation learning approach that generates multiple candidate actions and learns to score them. This architecture enables both fast inference and effective modeling of multimodal behaviors. We validate our approach on two real-world tasks: dishwasher loading and whole-body loco-manipulation for whiteboard wiping. Experiments show that Choice Policy significantly outperforms diffusion policies and standard behavior cloning. Furthermore, our results indicate that hand-eye coordination is critical for success in long-horizon tasks. Our work demonstrates a practical path toward scalable data collection and learning for coordinated humanoid manipulation in unstructured environments.

Figures (5)

• Figure 1: An overview of our modular teleoperation interface. Control is decomposed into four modules: arm control (end-effector tracking), hand control (trigger button for power/precision grasps and joystick for thumb control), head control (hand-eye tracking of the left or right hand), and locomotion control (omnidirectional walking). A single joystick is shared between thumb control and locomotion; pressing the joystick switches between the two modes. The figure illustrates two humanoid platforms: one mounted on a fixed shelf for manipulation tasks without locomotion, and another with full walking capability. This modular design simplifies data collection while preserving whole-body coordination across arms, hands, head, and legs.
• Figure 2: Policy architectures for imitation learning. (a) Diffusion Policy models multi-modality but requires $K$ iterative sampling steps, which makes inference slow. (b) Standard Behavior Cloning enables fast inference with a single forward pass but struggles to capture multimodal behaviors. (c) Choice Policy (ours) combines the strengths of both: it generates $K$ candidate actions in a single forward pass and uses a learned score to select the best one. This design achieves fast inference while effectively handling multi-modality.
• Figure 3: PyTorch pseudocode for the Choice Policy training and inference procedure. We include this snippet to highlight that our method is simple to implement; it requires only a few lines of code to combine winner-takes-all action learning with score regression. This ease of implementation makes the approach readily reproducible.
• Figure 4: Camera view comparison: (A) without hand-eye coordination vs. (B) with hand-eye coordination. The figure shows head and wrist camera perspectives during the insertion stage. During insertion, the head camera with hand-eye coordination provides a clear view of the dishrack. By contrast, without coordination, the view is occluded, which makes it difficult to verify whether the plate has been successfully inserted.
• Figure 5: Visualization of policy specialization and switching. We visualize the selection of action proposals during a complete rollout of the dishwasher loading task. Each blue row corresponds to one of the $K=5$ action proposals (Choice IDs 0-4), while the columns represent different task phases. The figure illustrates how different proposal heads specialize in specific sub-skills. At inference time, the score prediction network selects the most suitable expert head for the current phase, which enables the policy to maintain high precision across the entire long-horizon task.