AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control

TL;DR

This work introduces Adaptive Motion Optimization (AMO), a framework that fuses trajectory optimization with sim-to-real reinforcement learning to achieve real-time, hyper-dexterous humanoid whole-body control. AMO addresses embodiment and distribution biases by coupling an Adaptation Module trained on a novel AMO dataset with a generalizable lower-policy and an upper policy that operates via teleoperation or autonomous imitation. The approach demonstrates expanded manipulation workspace, improved stability, and robust OOD performance in both simulation and real-world 29-DoF humanoids, including autonomous task execution after demonstrations. Overall, AMO offers a scalable, robust pathway for dexterous loco-manipulation on high-DoF humanoids, with potential for broader autonomous and teleoperation applications on real hardware.

Abstract

Humanoid robots derive much of their dexterity from hyper-dexterous whole-body movements, enabling tasks that require a large operational workspace: such as picking objects off the ground. However, achieving these capabilities on real humanoids remains challenging due to their high degrees of freedom (DoF) and nonlinear dynamics. We propose Adaptive Motion Optimization (AMO), a framework that integrates sim-to-real reinforcement learning (RL) with trajectory optimization for real-time, adaptive whole-body control. To mitigate distribution bias in motion imitation RL, we construct a hybrid AMO dataset and train a network capable of robust, on-demand adaptation to potentially O.O.D. commands. We validate AMO in simulation and on a 29-DoF Unitree G1 humanoid robot, demonstrating superior stability and an expanded workspace compared to strong baselines. Finally, we show that AMO's consistent performance supports autonomous task execution via imitation learning, underscoring the system's versatility and robustness.

Figures (7)

• Figure 1: System overview. The system is decomposed into four stages: 1. AMO module training by collecting AMO dataset using trajectory optimization; 2. RL policy training by teacher-student distillation in simulation; 3. real robot teleoperation by IK and retargeting; 4. real robot autonomous policy training by imitation learning (IL) with a transformer.
• Figure 2: Teleoperation system overview. The operator provides three end-effector targets: head, left wrist, and right wrist poses. A multi-target IK computes upper goals and intermediate goals by matching three weighted targets simultaneously. The intermediate goals ($\mathbf{rpy}, h$) are fed to AMO and converted to lower goals.
• Figure 3: Comparison of torso orientation ranges.
• Figure 4: Evaluation of in-distribution (I.D.) and out-of-distribution (O.O.D.) tracking results. Each figure shows the target vs. the actual commanded direction. The white area indicates I.D., meaning the command is used in both trajectory optimization and RL training. The grey area indicates O.O.D., meaning the command is not used in either trajectory optimization or RL training. The red and blue curves represent w.o. AMO and AMO, respectively.
• Figure 5: Autonomous tasks performed in the real-world setting. For each task, we collect 50 episodes using the teleoperation system and train an ACT to complete it autonomously.