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

Technical Summary of MOTIONGRPO

Problem Statement
Recovering full-body 3D human motion from Head-Mounted Device (HMD) signals, such as those from Project Aria, remains a critical challenge for VR/AR applications. Unlike third-person capture, egocentric settings suffer from severe occlusion where the user's body is largely unobserved. Existing diffusion-based methods address this by learning motion priors conditioned on head trajectories. However, these methods often rely on global distribution matching, leading to local joint reconstruction errors, spatial misalignment, and visual artifacts such as foot skating, motion jitter, and ground penetration. Furthermore, applying Reinforcement Learning (RL) directly to diffusion-based motion recovery faces a specific bottleneck: the "low intra-group diversity" problem. Because motion recovery is heavily constrained by deterministic head signals, sampled outputs within a group exhibit minimal variance. This lack of diversity causes the standard deviation in Group Relative Policy Optimization (GRPO) advantage calculations to approach zero, resulting in vanishing gradients and unstable training.

Methodology
The authors propose MOTIONGRPO, a framework that leverages RL post-training to inject fine-grained guidance into the diffusion sampling process. The methodology is structured around three core components:

1. MDP Formulation and GRPO: The diffusion sampling process is modeled as a Markov Decision Process (MDP). The policy is optimized using Group Relative Policy Optimization (GRPO), which estimates value baselines directly from group-wise scores without requiring a separate value network. The policy $\pi_\theta$ corresponds to the reverse transition distribution of the diffusion model.
2. Hybrid Reward Mechanism: To guide the optimization, a hybrid reward system is designed:
   • Global Visual Reward: A trajectory-conditioned perceptual model evaluates visual plausibility. This model is trained via online contrastive learning, actively synthesizing hard negative samples (generated motions paired with head trajectories) to distinguish natural motions from artifacts like foot skating and jitter.
   • Local Joint Reward: Explicit sub-rewards enforce precision on joint positions, rotations, and velocities. These include pose rewards computed with and without per-frame similarity transform alignment, local rotation discrepancies, and velocity differences.
3. Noise-Injection Strategy: To address the low intra-group diversity bottleneck, the authors introduce a noise-injection strategy. Temporally smoothed Perlin noise is added to the translation component of the input head trajectory conditions. This simulates pseudo out-of-distribution inputs, forcing the diffusion policy to face slightly shifted states. This increases model uncertainty and output diversity within the sampled group, ensuring a non-trivial standard deviation for the GRPO advantage calculation and stabilizing the learning process.

Key Contributions

• MOTIONGRPO Framework: The introduction of an RL-based post-training framework for egocentric motion recovery that optimizes a hybrid reward combining a learned perceptual model for global plausibility and explicit objectives for precise joint alignment.
• Identification and Mitigation of Low Diversity: The identification of "low intra-group diversity" as a fundamental bottleneck for applying GRPO to deterministic motion recovery tasks. The authors propose a temporally smoothed noise-injection strategy to increase output diversity and prevent vanishing gradients.
• State-of-the-Art Performance: Extensive experiments demonstrating that the framework achieves superior performance in both joint accuracy and visual fidelity compared to existing baselines.

Experimental Results
The method was evaluated on the AMASS and RICH benchmarks, with additional qualitative validation on the Aria Digital Twins (ADT) dataset.

• Quantitative Performance: MOTIONGRPO achieved state-of-the-art results, significantly outperforming competitive baselines like EgoAllo. On the AMASS dataset, it reduced the Mean Per-Joint Position Error (MPJPE) from 124.985 mm to 114.207 mm and the Procrustes-Aligned MPJPE (PA-MPJPE) from 103.958 mm to 95.512 mm. Similar improvements were observed on the RICH dataset.
• Visual Quality: The framework effectively mitigated high-frequency artifacts. Metrics for Jitter, Foot Skating, and Ground Penetration showed consistent improvements over baselines, confirming the efficacy of the hybrid reward mechanism.
• Ablation Studies:
  • Vanilla GRPO with joint rewards improved local precision but required the visual reward to address global physical plausibility.
  • The Perlin noise injection strategy was critical; without it, group diversity collapsed, and performance gains were limited.
  • The use of synthesized hard negative samples significantly improved the perceptual model's ability to discriminate between natural and flawed motions compared to simple noise injection on ground truth.
• Efficiency: As a post-training strategy, MOTIONGRPO imposes negligible computational overhead during inference, maintaining inference speeds and VRAM usage comparable to the baseline diffusion models.

Significance
The paper claims that MOTIONGRPO represents a significant advancement by bridging the gap between generative diffusion models and the precise geometric constraints required for motion recovery. By adapting GRPO with a novel noise-injection strategy, the framework overcomes the limitations of standard distribution matching and the vanishing gradient problem inherent in deterministic recovery tasks. The result is a system capable of recovering high-fidelity full-body motion with superior visual plausibility and joint accuracy, offering a robust solution for egocentric motion recovery in VR/AR applications.