ICML 2026

MotionGRPO
Overcoming Low Intra-Group Diversity in GRPO-Based Egocentric Motion Recovery

A reinforcement learning framework that post-trains diffusion models to recover high-fidelity full-body 3D human motion from head-mounted device signals.

Nanjie Yao · Junlong Ren · Wenhao Shen · Hao Wang

The Hong Kong University of Science and Technology (Guangzhou) · Nanyang Technological University

Code arXiv BibTeX
Task Overview

Egocentric Motion Recovery

Given head trajectory signals and optional egocentric images from a head-mounted device, MotionGRPO recovers high-fidelity full-body 3D human motion in 3D scenes.

MotionGRPO Task Overview
Abstract

Overview

Recovering full-body 3D human motion from sparse head-mounted device signals.

Abstract

This paper studies full-body 3D human motion recovery from head-mounted device signals. Existing diffusion-based methods often rely on global distribution matching, leading to local joint reconstruction errors. We propose MotionGRPO, a novel framework leveraging reinforcement learning post-training to inject fine-grained guidance into the diffusion process. Technically, we model diffusion sampling as a Markov decision process optimized via Group Relative Policy Optimization (GRPO). To this end, we introduce a hybrid reward mechanism that combines a learned conditioned perceptual model for global visual plausibility and explicit constraints for local joint precision. Our key technical insight is that policy optimization in diffusion-based recovery suffers from vanishing gradients due to limited intra-group sample diversity. To address this, we further introduce a noise-injection strategy that explicitly increases sample variance and stabilizes learning. Extensive experiments demonstrate that MotionGRPO achieves state-of-the-art performance with superior visual fidelity.

🎯

RL-Based Motion Recovery Framework

We propose MotionGRPO, which optimizes a hybrid reward combining a trajectory-conditioned perceptual model for global plausibility and fine-grained objectives for precise joint alignment.

🎲

Noise-Injection Strategy

We identify the "low intra-group diversity" bottleneck in GRPO and introduce temporally smoothed noise injection to increase output diversity and stabilize training.

🏆

State-of-the-Art Performance

Extensive experiments on AMASS and RICH benchmarks show superior visual fidelity and strong generalization to real-world scenarios.

Method

How It Works

MotionGRPO formulates diffusion sampling as a Markov Decision Process (MDP), optimized via GRPO with a hybrid reward mechanism. A noise-injection strategy overcomes the low intra-group diversity bottleneck.

MotionGRPO Framework Overview
Hybrid Reward

Perceptual + Geometric Guidance

A trajectory-conditioned perceptual model trained with online contrastive learning detects visual artifacts like foot skating and motion jitter. Combined with explicit sub-rewards on joint positions, rotations, and velocities, the hybrid reward guides the diffusion model toward both visually plausible and geometrically accurate motions.

Noise Injection

Overcoming Low Intra-Group Diversity

Directly applying GRPO to motion recovery suffers from vanishing gradients because strong conditioning from head signals constrains the output space. Our temporally smoothed noise-injection strategy simulates out-of-distribution inputs, increasing model uncertainty and output diversity for effective GRPO optimization.

Results

Experimental Results

MotionGRPO achieves state-of-the-art performance on both AMASS and RICH benchmarks.

Method AMASS Dataset RICH Dataset
Joint Accuracy Visual Quality Joint Accuracy Visual Quality
MPJPE
(mm) ↓		 PA-MPJPE
(mm) ↓		 MPJVE
(mm/s) ↓		 MPJRE
(°) ↓		 Jitter
↓		 GP
(m) ↓		 FS
(m) ↓		 MPJPE
(mm) ↓		 PA-MPJPE
(mm) ↓		 MPJVE
(mm/s) ↓		 MPJRE
(°) ↓		 Jitter
↓		 GP
(m) ↓		 FS
(m) ↓
EgoEgo 177.231152.125588.6619.457 2.6431.3311.241 221.450196.223572.33113.312 5.1874.3571.021
EgoAllo 124.985103.958553.2218.777 2.3941.1431.290 192.686172.724506.99212.734 4.1354.1451.094
EgoAlloℵ 121.651101.034483.4718.728 1.4551.0990.479 190.000169.838407.62812.638 1.8804.4380.223
MotionGRPO 114.20795.512531.2178.413 2.0000.9011.169 187.223169.146477.34411.944 3.6853.1611.008
MotionGRPOℵ 111.77693.702461.7028.330 1.3090.9630.399 184.992167.032378.42311.886 1.6143.1560.199

Bold: best, Underline: second best. ℵ denotes test-time post-processing. ↓ lower is better.

Qualitative comparison results

Qualitative comparison on AMASS and ADT datasets. MotionGRPO produces more accurate joint alignment and fewer visual artifacts, with strong generalization to real-world scenarios.

Method AMASS Dataset RICH Dataset
MPJPE ↓PA-MPJPE ↓MPJVE ↓MPJRE ↓ Jitter ↓GP ↓FS ↓ MPJPE ↓PA-MPJPE ↓MPJVE ↓MPJRE ↓ Jitter ↓GP ↓FS ↓
Baseline 124.985103.958553.2218.777 2.3941.1431.290 192.686172.724506.99212.734 4.1354.1451.094
+ Vanilla GRPO 117.41897.945543.0128.403 2.0840.9991.272 190.248171.223494.12512.369 3.7933.6331.076
+ Visual Reward 116.54996.729546.1288.427 2.0140.8861.221 189.103170.002497.48012.268 3.6843.2251.044
+ Perlin Noise 114.20795.512531.2178.413 2.0000.9011.169 187.223169.146477.34411.944 3.6853.1611.008

Ablation study validating each component: Vanilla GRPO → Visual Reward → Perlin Noise Injection. Each addition progressively improves both joint accuracy and visual quality.

Citation

BibTeX

If you find this work useful, please cite our paper.

@inproceedings{ yao2026motiongrpo, title={Motion{GRPO}: Overcoming Low Intra-Group Diversity in {GRPO}-Based Egocentric Motion Recovery}, author={Nanjie Yao and Junlong Ren and Wenhao Shen and Hao Wang}, booktitle={Forty-third International Conference on Machine Learning}, year={2026}, }
arXiv GitHub Contact