HumanoidMimicGen
Data Generation for Loco-Manipulation via Whole-Body Planning and Adaptation

Kevin Lin^*, Ajay Mandlekar^*, Caelan Reed Garrett^*, Nikita Chernyadev, Yu Fang,
Runyu Ding, Yuqi Xie, Justin Tran, Linxi Fan^†, Yuke Zhu^†

NVIDIA · The University of Texas at Austin · ^*Equal contribution · ^†Project leads

Generate thousands of humanoid loco-manipulation demonstrations from a single teleoperated demonstration.

Paper arXiv Summary

Real-World Results

HumanoidMimicGen-generated data improves real world policy performance.

Co-trained policyHumanoidMimicGen data + real robot dataSuccess

Real-only policyReal robot data onlyFailure

Autonomous | 1x speed

BoxToCart

Abstract

From a few demos to thousands.

Imitation learning can train humanoids to walk and manipulate, but teleoperated demos are expensive. Existing data-generation methods work well for arms, but transfer poorly to humanoids with coupled arms, legs, and torso control.

HumanoidMimicGen generates humanoid loco-manipulation data by adapting contact-rich whole-body skills from a few source demos to new states and object poses. It interleaves single- and dual-arm skills with locomotion and manipulation planning to produce stable trajectories in varied layouts.

We evaluate on a new simulated G1 loco-manipulation benchmark and study how data generation and policy design affect success. We also show that co-training policies on HumanoidMimicGen-generated data with a small real-world dataset outperforms those trained only on the real-world data by 20%.

Method

Skill adaptation plus whole-body planning.

HumanoidMimicGen starts from a single source demonstration with per-arm skill annotations and ordering constraints. These annotations define when the base and arms move, and the target end-effector pose for each skill.

Whole-body planning builds a locomotion plan and per-arm plans to reach those targets. The upper body uses joint-space control; the lower body follows pelvis-velocity commands from a learned locomotion policy.

HumanoidMimicGen generates data by interleaving locomotion planning, manipulation planning, and demonstrated skill adaptation

Locomotion
Planning

Manipulation
Planning

Skill
Adaptation

Planning generates demos across new object poses and obstacle layouts with the same source demonstration.

Drill Lift

Obstacle-Aware Pick Drill

Contributions

Method, benchmark, and hardware results.

HumanoidMimicGen Algorithm

A data-generation method that turns a few humanoid teleoperation demos into many planned loco-manipulation trajectories.

G1 Loco-Manipulation Benchmark

Nine simulated G1 loco-manipulation tasks for comparing data-generation and policy-learning methods.

Policy Learning Analysis

Experiments on noise, data-generation methods, and model architectures for policy learning.

Sim-and-Real Co-Training

Hardware experiments showing that sim + real co-training improves real-world success rates.

Benchmark

Nine industrial humanoid tasks.

Object lifting, pushing, placing, obstacle navigation, and shelf interaction on a humanoid robot.

Box Lift Floor task snapshots — 01 Box Lift Floor Grasp the box from the floor and lift to a target height.

Push Button task snapshots — 02 Push Button Approach an industrial panel and press its button.

Box Lift task snapshots — 03 Box Lift Approach the table, grasp the box, and lift to a target height.

Push Shelf Forward task snapshots — 04 Push Shelf Forward Push the shelving cart into a marked target zone.

Drill Lift task snapshots — 05 Drill Lift Approach a table, grasp a drill, and lift to a target height.

Drill PnP task snapshots — 06 Drill PnP Pick a drill from one table and place it on a second table.

Box Table To Shelf task snapshots — 07 Box Table To Shelf Transfer a box from a table into a shelf.

Pick Drill From Holder task snapshots — 08 Pick Drill From Holder Extract a drill from a holder and lift to a target height.

Obstacle-Aware Pick Drill task snapshots — 09 Obstacle-Aware Pick Drill Navigate around a blocking shelf, then grasp and lift the drill.

Simulation Results

89% average success.

HumanoidMimicGen reaches the best average success across all nine tasks.

Task	1 Human Demo	100 Human Demos	DexMimicGen+ (1k Generated)	Ours (1k Generated)
Box Lift Floor	14%	83%	87%	97%
Push Button	18%	82%	26%	92%
Box Lift	95%	95%	68%	100%
Push Shelf Forward	70%	90%	35%	100%
Drill Lift	20%	30%	13%	100%
Drill PnP	8%	20%	13%	70%
Box Table To Shelf	4%	28%	17%	53%
Pick Drill From Holder	0%	7%	36%	100%
Obstacle-Aware Pick Drill	4%	0%	0%	87%
Average	26%	48%	33%	89%

Success rates for VLAs finetuned on 1 human demo, 100 human demos, 1,000 demos generated by DexMimicGen+, and 1,000 demos generated by our method, HumanoidMimicGen. For each task, both DexMimicGen+ and HumanoidMimicGen use the same single source demonstration to generate the 1,000 synthetic demonstrations.