VLA Gallery / All models / Humanoid whole-body controllers

AMO

AMO (Adaptive Motion Optimization)

UC San Diego · 2025-05

VAHumanoid whole-body controllersNot disclosedRL policy

Adapts to torso commands using hybrid data generated by trajectory optimization (Crocoddyl BoxFDDP), performing whole-body manipulation with very high DoF.

Paper

AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body ControlarXiv 2505.03738

Overview

Category

Humanoid whole-body controllers

Type

VA

Released

2025-05

Organization

UC San Diego

Open weights

Partial

Parameters

Not disclosed

Architecture

Backbone LLM

None

Vision encoder

—

Action head

AMO 3-layer MLP (RL policy)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

AMO module: 3-layer MLP; PPO teacher → supervised student; autonomous upper-body policy is ACT (Transformer) + DINOv2 encoder

Training

Pretraining accelerators

Not disclosed (IsaacGym massively parallel sim)

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed (env count Not disclosed)

GPU/TPU-hours

Not disclosed

Fine-tuning

N/A; imitation learning uses 50 episodes per task

Data type

Motion capture / simulation·synthetic / real-robot teleoperation

Data scale

AMASS upper-body motion + random torso commands (scale not disclosed); 50 episodes per imitation learning task

Main datasets

AMASS

Tasks & evaluation

Torso tracking; paper-bag pickup 8~10/10, plastic-bottle tossing 10/10, basket pickup (qualitative evaluation)

Robot

Unitree G1 (29-DoF) + Dex3-1 hands, 3-DoF active head, ZED Mini

Inference

Inference device & speed

Jetson Orin NX 50 Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Onboard (Jetson Orin NX)

Reinforcement learning & base

Reinforcement learning

Used (RL controller training) · AMO module is PPO teacher → supervised student. Autonomous upper-body policy (ACT) uses imitation learning

Base control

Humanoid legged (Commands/latents → RL locomotion controller) · Autonomous ACT policy outputs upper-body joints + lower-body velocity/pose/height commands → AMO RL lower-body policy generates leg joint targets (walking, turning)

Notes

Key features

Adapts to torso commands using hybrid data generated by trajectory optimization (Crocoddyl BoxFDDP), performing whole-body manipulation with very high DoF. Converts teleoperation·imitation learning commands into whole-body joint targets.

Notes

RSS 2025. The repo provides only a MuJoCo interactive demo (play_amo.py), so weights are 'Partial'. Training compute·parameters NOT FOUND in verification either.

Verification

Confirmed