Research Analyzer
← Back ICRA 2026

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

PDF

AI summary

Key figure (auto-extracted from paper)
Using a human motion diffusion prior to guide reinforcement learning enables autonomous whole-body humanoid skills that direct RL cannot discover, while naturally improving sim-to-real transfer.
Humanoid control Diffusion models Reinforcement learning Whole-body manipulation Sim-to-real transfer Motion priors

Problem

Training humanoids for whole-body interaction tasks is hindered by the scarcity of teleoperation data, the complexity of coordinating multiple timescales, and the difficulty of RL exploration in high-dimensional spaces. Direct RL often fails or produces unnatural motions that do not transfer well to real robots.

Approach

The method generates human-like motion plans using a text- and guidance-conditioned diffusion model trained on abundant human motion data, then trains a reinforcement learning policy in simulation to follow these plans while completing specific interaction tasks, before deploying the policy to a real robot.

Key results

  • Enables autonomous whole-body loco-manipulation without teleoperation data
  • RL policy discovers complex solutions unattainable by direct exploration
  • Achieves successful sim-to-real transfer on a Unitree G1 robot
  • Supports diverse tasks including drawer opening, bimanual picking, and button pressing

Why it matters

It provides a scalable, data-efficient pathway for training complex whole-body manipulation skills in humanoids, accelerating the development of practical humanoid assistants.

Abstract

We introduce DreamControl, a novel method- ology 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: genrobo.github.io/DreamControl/ Appendix: genrobo.github.io/DreamControl/Appendix.pdf

Index terms

Whole-Body Motion Planning and Control Humanoid Robot Systems Reinforcement Learning

Related papers