• [2026-07] 🎉 MotionGRPO has been accepted to ICML 2026!
• [2026-05] 📄 Preprint available on arXiv.
• 🔜 [2026-07] We are actively extending this work to hand motion recovery. The enhanced codebase and the extended preprint will be released in October 2026. Stay tuned!

Recovering full-body 3D human motion from head-mounted device signals remains a critical challenge for VR/AR applications. Existing diffusion-based methods often rely on global distribution matching, leading to local joint reconstruction errors and visual artifacts such as foot skating and ground penetration.

MotionGRPO addresses these issues by leveraging Reinforcement Learning (RL) post-training to inject fine-grained guidance into the diffusion process. We model diffusion sampling as a Markov Decision Process (MDP) and optimize it via Group Relative Policy Optimization (GRPO) with a hybrid reward mechanism. A trajectory-conditioned perceptual model ensures global visual plausibility, while explicit sub-rewards target joint positions, rotations, and velocities for local precision. To overcome the low intra-group diversity bottleneck that causes vanishing gradients in GRPO, we introduce a temporally smoothed noise-injection strategy that increases output variance and stabilizes training.

Extensive experiments on AMASS and RICH benchmarks demonstrate that MotionGRPO achieves state-of-the-art performance with superior visual fidelity.

• 🏠 Project Page: https://3dagentworld.github.io/MotionGRPO
• 📄 arXiv: https://arxiv.org/abs/2605.05680
• 💻 Code: Coming soon — the official implementation and an enhanced version (with hand recovery extension) are expected to be released in October 2026.

MotionGRPO formulates diffusion sampling as an MDP and optimizes it via GRPO with a hybrid reward mechanism.

1. 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.

2. 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.

3. State-of-the-Art Performance — Extensive experiments on AMASS and RICH benchmarks show that MotionGRPO achieves superior visual fidelity and strong generalization to real-world scenarios.

MotionGRPO achieves state-of-the-art results on both AMASS and RICH benchmarks across joint accuracy and visual quality metrics.

See the paper and project page for the full quantitative and qualitative results.

If you find this work useful, please cite:

@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},
}

For questions or collaborations, please contact Hao Wang or open an issue.