Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DreamControl: Human-Inspired Whole-Body Humanoid Control for Scene Interaction via Guided Diffusion

Dvij Kalaria, Sudarshan S Harithas, Pushkal Katara, Sangkyung Kwak, Sarthak Bhagat, Shankar Sastry, Srinath Sridhar, Sai Vemprala, Ashish Kapoor, Jonathan Chung-Kuan Huang

TL;DR

DreamControl addresses the challenge of learning autonomous whole-body humanoid skills by fusing a diffusion-based human motion prior with reinforcement learning. A diffusion prior guided by text and spatiotemporal cues generates reference trajectories, which are then retargeted to a Unitree G1 and used to train a goal-conditioned RL policy to execute tasks in simulation and transfer to real hardware. The approach yields more natural, stable, and task-consistent motions than baselines, with strong sim2real performance and broad task coverage. By reducing reliance on teleoperation data and leveraging abundant human motion data, DreamControl offers a data-efficient route to scalable, scene-interacting humanoid control across diverse morphologies and tasks.

Abstract

We introduce DreamControl, a novel methodology for learning autonomous whole-body humanoid skills. DreamControl leverages the strengths of diffusion models and Reinforcement Learning (RL): our core innovation is the use of a diffusion prior trained on human motion data, which subsequently guides an RL policy in simulation to complete specific tasks of interest (e.g., opening a drawer or picking up an object). We demonstrate that this human motion-informed prior allows RL to discover solutions unattainable by direct RL, and that diffusion models inherently promote natural looking motions, aiding in sim-to-real transfer. We validate DreamControl's effectiveness on a Unitree G1 robot across a diverse set of challenging tasks involving simultaneous lower and upper body control and object interaction. Project website at https://genrobo.github.io/DreamControl/

DreamControl: Human-Inspired Whole-Body Humanoid Control for Scene Interaction via Guided Diffusion

TL;DR

DreamControl addresses the challenge of learning autonomous whole-body humanoid skills by fusing a diffusion-based human motion prior with reinforcement learning. A diffusion prior guided by text and spatiotemporal cues generates reference trajectories, which are then retargeted to a Unitree G1 and used to train a goal-conditioned RL policy to execute tasks in simulation and transfer to real hardware. The approach yields more natural, stable, and task-consistent motions than baselines, with strong sim2real performance and broad task coverage. By reducing reliance on teleoperation data and leveraging abundant human motion data, DreamControl offers a data-efficient route to scalable, scene-interacting humanoid control across diverse morphologies and tasks.

Abstract

We introduce DreamControl, a novel methodology for learning autonomous whole-body humanoid skills. DreamControl leverages the strengths of diffusion models and Reinforcement Learning (RL): our core innovation is the use of a diffusion prior trained on human motion data, which subsequently guides an RL policy in simulation to complete specific tasks of interest (e.g., opening a drawer or picking up an object). We demonstrate that this human motion-informed prior allows RL to discover solutions unattainable by direct RL, and that diffusion models inherently promote natural looking motions, aiding in sim-to-real transfer. We validate DreamControl's effectiveness on a Unitree G1 robot across a diverse set of challenging tasks involving simultaneous lower and upper body control and object interaction. Project website at https://genrobo.github.io/DreamControl/

Paper Structure

This paper contains 49 sections, 24 equations, 3 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Recent Advances in Manipulation
  4. RL controllers for legged robots
  5. Character Animation and Motion Models
  6. The DreamControl methodology for constructing humanoid skills
  7. Stage 1: Generating Reference Trajectory from a Human Motion Prior
  8. Stage 2: RL with Reference Trajectory
  9. Evaluation
  10. Tasks and Baselines
  11. Simulation Results
  12. Human-ness Comparison
  13. Sim2Real Deployment
  14. Discussion and Future Work
  15. Robot details
  16. ...and 34 more sections

Figures (3)

  • Figure 1: Unitree G1 humanoid performing various skills trained via DreamControl, including (1) opening a drawer, (2) bimanual pick (of a box), (3) ordinary pick and (4) pressing an elevator button.
  • Figure 2: DreamControl Overview: (A) we first generate text and spatiotemporally guided human motion trajectories using diffusion; (B) we train goal-conditioned RL policies to track these generated trajectories while completing some task of interest; (C) we deploy these policies to a real humanoid, leveraging off-the-shelf vision models to determine spatial guidance inputs for the RL policy.
  • Figure 3: Comparison of trajectories for the task of Jump. The top row shows results from the TaskOnly+ baseline, while the bottom row illustrates trajectories from the DreamControl. The yellow sphere depicts the spatial control point used to guide the trajectories.