Research map / Reinforcement learning / RL from feedback
MotionGRPO: Overcoming Low Intra-Group Diversity in GRPO-Based Egocentric Motion Recovery
Nanjie Yao, Junlong Ren, Wenhao Shen, Hao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add vanilla GRPO post-training with joint rewards (ablation) | EgoAllo diffusion baseline · AMASS | MPJPE (mm): 124.985 -> 117.418 | Improved |
| add visual reward from perceptual model (ablation) | Vanilla GRPO · AMASS | Ground penetration (m): 0.999 -> 0.886 | Improved |
| add visual reward (ablation) | Vanilla GRPO · AMASS | MPJPE (mm): 117.418 -> 116.549 | Improved |
| add Perlin noise injection (ablation) | Vanilla GRPO + Visual · AMASS | MPJPE (mm): 116.549 -> 114.207 | Improved |
| add Perlin noise injection (ablation) | Vanilla GRPO + Visual · RICH | MPJPE (mm): 189.103 -> 187.223 | Improved |
| add Perlin noise injection (ablation) | Vanilla GRPO + Visual · AMASS | Jitter: 2.014 -> 2.000 | No clear effect |
| increase Perlin noise intensity from 0 to 0.05 (ablation) | GRPO sampling · Training subset | diversity: 1.8827 -> 2.2048 | Improved |
| increase Perlin noise intensity from 0.05 to 0.1 (ablation) | GRPO sampling · Training subset | diversity: 2.2048 -> 3.1430 | Improved |
| hard-negative samples instead of GT-noise negatives (ablation) | Perceptual model · AMASS perceptual model | accuracy (%): 86.04 -> 97.68 | Improved |
| GRPO instead of direct fine-tuning | EgoAllo · AMASS | MPJPE (mm): 124.980 -> 114.207 | Improved |
| GRPO instead of DPO | EgoAllo · AMASS | MPJPE (mm): 118.043 -> 114.207 | Improved |
| DPO versus baseline on ground penetration | EgoAllo · AMASS | Ground penetration (m): 1.143 -> 1.530 | Hurt |
Closest papers by meaning
- REWIND: Real-Time Egocentric Whole-Body Motion Diffusion with Exemplar-Based Identity Conditioning (CVPR 2025 · similarity 0.85)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.85)
- Optimizing Diffusion Noise Can Serve As Universal Motion Priors (CVPR 2024 · similarity 0.85)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.85)
- PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
