🧑 Human Understanding

🧪 ICML2026 · 1 paper notes

📌 Same area in other venues: 📷 CVPR2026 (53) · 🔬 ICLR2026 (7) · 🤖 AAAI2026 (15) · 🧠 NeurIPS2025 (18) · 📹 ICCV2025 (38)

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

MotionGRPO reformulates egocentric full-body motion recovery from head-mounted devices as an MDP over diffusion sampling, employing GRPO with a "trajectory-conditioned perceptual model + 4 joint-level sub-rewards" hybrid reward for post-training. It identifies the critical bottleneck of "overly strong input conditions leading to nearly identical intra-group samples and vanishing advantage variance," and restores intra-group diversity by injecting Perlin noise into the conditions. On AMASS/RICH, it reduces MPJPE from EgoAllo's 124.985 mm to 114.207 mm.