Source-linked AI summary
AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control
Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Ri-Zhao Qiu, Xiaolong Wang
TL;DR
Humanoid whole-body control must handle high-DoF, nonlinear dynamics while supporting dexterous motions beyond motion-imitation training distributions. AMO combines trajectory optimization, hybrid-data training, and sim-to-real RL to adapt whole-body control in real time. Experiments report expanded workspace, stronger torso stability and tracking, O.O.D. command generalization, and autonomous basket picking, while decoupled arm-base control limits whole-body coordination in highly dynamic scenarios.
Problem
High-DoF, nonlinear humanoid dynamics and biased motion-imitation data limit real-time whole-body dexterity and adaptation to O.O.D. commands.
Method
AMO combines trajectory optimization with hierarchical sim-to-real RL, using hybrid motion data and an adaptive module that converts torso commands into lower-body references.
Results
AMO expands torso motion, improves roll and pitch tracking, generalizes beyond training ranges, and supports autonomous real-world basket picking.
Takeaways & Limitations
AMO provides real-time humanoid control with expanded operational workspace and adaptive behavior for teleoperation and autonomous loco-manipulation.
Takeaways & Limitations
Decoupled arm control is independent of the robot base state, limiting whole-body coordination in highly dynamic scenarios.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
AMO addresses the difficulty of real-time hyper-dexterous humanoid control by combining hybrid motion synthesis with adaptive policy training. The framework targets broader workspace, O.O.D. commands, and autonomous whole-body tasks through simulation, real-world, teleoperation, and imitation-learning evaluations.
- Humanoid whole-body control is difficult because 29-DoF robots have high-dimensional, nonlinear, contact-rich dynamics that challenge real-time model-based control.Traditional methods also require precise models, substantial computation, and reduced-order approximations.
- Motion-capture datasets are biased toward bipedal locomotion, while trajectory-optimization methods use limited motion primitives and are computationally inefficient for real-time generalization.These limitations restrict coordinated arm-torso movements and adaptation to unstructured inputs.
- AMO fuses motion-capture arm trajectories with sampled torso orientations, then uses dynamics-aware trajectory optimization to construct whole-body reference motions.The resulting hybrid dataset is designed for dexterous loco-manipulation and addresses motion-distribution bias.
- The AMO network learns continuous command-to-motion mappings for interpolation across continuous inputs and O.O.D. teleoperation commands.This replaces discrete lookup-table behavior with real-time adaptive responses.
- The paper reports expanded humanoid workspace, O.O.D. performance, and a workspace large enough to enable picking objects from the ground.The introduction characterizes AMO as a real-time controller using sparse task-space targets.
- AMO is evaluated through simulation, real-robot teleoperation, and autonomous imitation learning, including a basket-picking task requiring crouching and torso adjustment.The system overview separates AMO training, simulated RL distillation, teleoperation, and autonomous transformer-based imitation learning.
III. ADAPTIVE MOTION OPTIMIZATION
AMO is organized as a hierarchical whole-body-control framework for both teleoperation and autonomous operation. Teleoperation conditions the policy on human goals, whereas the autonomous policy generates commands from observations alone while retaining the same upper/lower-body hierarchy.
- AMO addresses humanoid whole-body control in two settings: teleoperation and autonomous operation.
- The teleoperation policy maps goals and observations to joint actions, while the autonomous policy maps observations directly to actions without human input.Teleoperation uses a goal-conditioned policy; autonomous control removes the human goal input.
- Teleoperation goals include head, left-hand, and right-hand poses plus base velocity, while observations combine stereo images and proprioception.Actions contain upper- and lower-body joint-angle commands.
- The upper policy outputs upper-body actions and intermediate torso-orientation and height commands, which the lower policy uses with velocity commands and proprioception.This hierarchical decomposition exposes rpy and h as intermediate lower-body controls.
- The autonomous policy preserves the same hierarchy and lower policy while its upper policy independently generates upper actions, velocity, and intermediate whole-body commands.
B. Adaptation Module Pre-Training
The AMO adaptation module converts torso orientation and height commands into lower-body joint references. It is trained from trajectory-optimized, dynamically feasible examples generated from hybrid upper-body motions and sampled torso commands.
- Whole-body torso and height tracking is difficult because standard locomotion sampling and foot-tracking heuristics do not provide guidance for coordinated whole-body control.Human-reference strategies also do not directly connect human poses with whole-body control.
- AMO maps upper-body joint inputs and torso commands to reference lower-body joint angles for explicit tracking by the lower policy.The module receives rpy and h and outputs references for all lower-body actuators.
- The AMO dataset is collected by sampling upper-body motions and torso commands, then solving a multi-contact optimal-control problem for lower-body joint references.Trajectory optimization tracks torso objectives while maintaining stable CoM and satisfying wrench-cone constraints.
- The optimization cost includes state and control regularization, torso-orientation and height tracking, and center-of-mass regularization for balance.
- A three-layer MLP is trained on these optimized data to convert torso commands into reference lower poses and remains frozen during later lower-policy training.The references are generated with control-limited feasibility-driven differential dynamic programming through Crocoddyl.
C. Lower Policy Training
The lower policy is trained in simulation with proprioceptive observations, using privileged-information teacher supervision and AMO-generated lower-body references to track locomotion and whole-body commands.
- Massively parallel IsaacGym simulation trains the lower policy to track intermediate commands g′ and base velocity v from proprioceptive observations.
- The lower policy observes base orientation, angular velocity, positions, velocities, position targets, upper-body actuator states, and gait phase.
- The lower action space q_lower ∈ R15 contains 12 leg-joint position targets and 3 waist-motor position targets.
- A teacher policy uses proprioceptive and privileged observations, including ground-truth base velocity, torso orientation, and base height.
- A student policy is distilled from the teacher using only real-world-accessible information, augmented with a 25-step proprioceptive history.
D. Teleoperation Upper Policy Implementation
The teleoperation upper policy uses retargeting and multi-target weighted inverse kinematics to coordinate head, wrist, torso-orientation, and height objectives across the upper body.
- Hand movements are generated by retargeting, while other teleoperation control signals are computed through inverse kinematics for manipulation precision.
- Multi-target weighted IK minimizes 6D distances to the head, left wrist, and right wrist simultaneously.
- The optimization variable includes all actuated upper-body DoFs and intermediate torso-orientation rpy and height h commands.
- Higher weights on torso and height terms than upper-body posture costs encourage upper-body actuators to handle simpler tasks first.
E. Autonomous Upper Policy Training
The autonomous upper policy is learned from teleoperated demonstrations and produces upper-body actions plus lower-policy commands, while evaluation compares tracking errors and torso control ranges.
- Autonomous upper policy training: Human operators teleoperate the robot with the goal-conditioned policy, recording observations and actions as demonstrations for imitation learning.
- Evaluation: Table II reports tracking errors averaged over 4096 environments and 500 steps, while Table III reports maximum torso ranges for in-distribution yaw, pitch, and roll commands.
- Autonomous upper policy training: The policy uses stereo images encoded by DINOv2 together with upper-body proprioception and the last commands sent to the lower policy.
- Autonomous upper policy training: Its output comprises upper-body joint angles and intermediate control signals for the lower policy.
- Evaluation: The experiments assess locomotion and torso-command tracking, comparisons with other whole-body-control strategies, and real-world performance on a 29-DoF Unitree G1 platform.
A. How well does AMO perform on tracking locomotion
AMO is evaluated against baselines on torso, height, and velocity tracking, with results emphasizing improved roll and pitch tracking, height tracking, and robustness to locomotion-related limitations.
- AMO is compared with w/o AMO, w/o priv, and w rand arms baselines using tracking-error metrics for torso orientation, height, and velocity.
- Torso orientation tracking: AMO achieves superior torso tracking in roll and pitch, with the most notable improvement occurring in pitch tracking.
- Torso orientation tracking: w rand arms has the lowest yaw tracking error, while w/o AMO struggles with yaw tracking; yaw error may not fully represent adaptive pose stability.
- Height tracking: AMO achieves the lowest height tracking error, whereas w/o AMO barely tracks height commands because it lacks reference information for learning the command-to-motor transformation.
- Linear velocity tracking: AMO maintains stable locomotion with low linear-velocity tracking error despite generating references under a double-support stance that omits foot-swing pose variations.
B. How does AMO compare to other WBC strategies?
AMO expands torso-control range and improves stability compared with waist-only, motion-tracking, and non-AMO baselines. It also tracks commands outside training distributions and supports autonomous whole-body tasks.
- Baseline comparisons: AMO is compared against w/o AMO, Waist Tracking, and ExBody2 for torso-control range.Table III provides quantitative measurements, while Fig. 4 shows qualitative differences.
- Baseline comparisons: ±0.52 radians is Unitree G1’s waist pitch positional limit, restricting Waist Tracking because it uses only waist motors.AMO instead uses the lower body to expand torso motion.
- AMO performance: AMO expands torso pitch, roll, and yaw control while dynamically adapting lower-body posture for stability.Examples include bending the upper body flat and slightly bending one leg to tilt the pelvis.
- O.O.D. tracking: AMO tracks out-of-distribution torso yaw up to ±2 despite training only to ±1.57, and tracks height down to 0.4m from a 0.5m–0.8m training range.w/o AMO fails on several O.O.D. pitch, yaw, and height commands.
- Autonomous tasks: AMO supports autonomous paper-bag picking, trash-bottle throwing, and long-horizon basket picking through imitation learning.Basket Picking requires crouching, torso adjustment, walking, and coordinated placement on an eye-level shelf.
- Autonomous tasks: Whole-body coordination in Basket Picking combines knee, waist pitch, and waist yaw trajectories across crouching, lateral grasping, walking, and shelf placement.The joint trajectories provide a case study of coordinated locomotion and manipulation.
V. CONCLUSIONS AND LIMITATIONS
AMO combines trajectory optimization and model-free RL for real-robot whole-body control, but its decoupled design limits coordination between arm control and the robot base.
- Conclusions: AMO integrates model-based trajectory optimization with model-free reinforcement learning for humanoid whole-body control.The framework is evaluated in simulation and on real-robot platforms.
- Limitations: The decoupled approach limits whole-body coordination because arm control is independent of the robot’s base state.The authors suggest incorporating base-state information into upper-limb joint-angle acquisition for balance-aware control.
APPENDIX A RETARGETING
The appendix describes motion retargeting through keypoint-vector matching and RL training through a staged curriculum that progressively adds delay, torso commands, and arm actions.
- Retargeting: The hand-retargeting objective minimizes distances between corresponding human and robot hand keypoint vectors.Robot keypoint vectors are derived from forward kinematics of the hand joint angles.
- Retargeting: The retargeting representation uses five vectors: three wrist-to-fingertip vectors and two thumb-to-fingertip vectors.Wrist vectors preserve overall pose, while thumb vectors support precise manipulation.
- Teacher training curriculum: RL training begins with locomotion and randomized target height, then adds delay, torso roll-pitch-yaw commands, and AMASS arm actions.The curriculum progressively trains the teacher policy for complex whole-body control.
- Teacher training curriculum: The curriculum randomizes target height, torso orientation, and arm joint positions after the curriculum begins, using ranges specified in Table V.Before that point, height is 0.8, torso orientation is zero, and arm positions remain at defaults.
- Curriculum schedules: Gait frequency and stance rate are defined as piecewise functions of global training step.The appendix associates these schedules with Table VI and denotes global step by s.
APPENDIX C DEPLOYMENT DETAILS
The teleoperation system and RL policy can run onboard a Jetson Orin NX while streaming images and human poses between VR devices and the robot.
- Onboard deployment: The full teleoperation system and RL policy run onboard a Jetson Orin NX at 50Hz inference frequency.Some experiments nevertheless use an external RTX 4090 PC for debugging.
- Onboard deployment: OpenTelevision streams images and human poses between the VR devices and the robot.This supports the onboard teleoperation setup.
- Onboard deployment: External-PC experiments are used for easier debugging, while deployment remains possible on the onboard Jetson platform.The passage distinguishes debugging hardware from the deployable system configuration.