Skeleton2Stage: Reward-Guided Fine-Tuning for Physically Plausible Dance Generation

1. Problem Statement

Current state-of-the-art (SOTA) dance generation methods primarily operate in the skeletal domain (representing motion via bone rotations). While these methods produce high-quality joint trajectories that align well with music, they often fail when visualized as a skinned human body mesh.

• The Skeleton-to-Mesh Gap: Skeletal representations ignore physical constraints inherent to a volumetric body mesh.
• Consequences: When converted to a mesh, generated dances frequently exhibit physically implausible artifacts, including:
  • Body Self-Penetration: Limbs or torso parts intersecting with each other.
  • Foot-Ground Contact (FGC) Anomalies: Feet floating above the ground or penetrating the floor.
  • Imbalanced Movements: Motions that defy gravity or friction.
• Limitation of Existing Solutions: Previous attempts to fix this often rely on post-hoc motion projection or optimization, which are computationally expensive, prone to failure (jittering), and often disable self-collision checks to ensure success, thereby failing to address full-body constraints.

2. Methodology: Skeleton2Stage

The authors propose Skeleton2Stage, a framework that bridges the skeleton-to-mesh gap by integrating Reinforcement Learning Fine-Tuning (RLFT) into diffusion models. The core idea is to use a physics simulator to generate rewards that guide the diffusion model toward physically plausible outputs.

A. Core Architecture

The method treats the diffusion denoising process as a Multi-Step Markov Decision Process (MDP). Instead of training from scratch, the authors take a pre-trained dance diffusion model (e.g., EDGE, POPDG) and fine-tune it using a policy gradient algorithm (REINFORCE) guided by specific rewards.

B. The Reward System

The framework employs a multi-faceted reward system to steer the model:

1. Imitation Reward (General Physical Plausibility):

   • Mechanism: A pre-trained imitation policy (trained on expert datasets like AMASS and AIST++) controls a character in a physics simulator (IsaacGym) using an SMPL mesh.
   • Logic: The policy attempts to mimic the motion generated by the diffusion model.
   • Reward Calculation: If the generated motion violates physics (e.g., self-penetration), the simulator's constraints prevent the policy from perfectly tracking it, resulting in a high tracking error. The reward is inversely proportional to this error.
   • Key Innovation: Unlike previous works, this setup enables self-collision checks in the simulator, forcing the diffusion model to learn to avoid body interpenetration to maximize the reward.

2. Foot-Ground Deviation (FGD) Reward & Guidance:

   • FGD Reward: Measures the distance between the lowest joint (feet) and the ground. It penalizes floating or deep penetration.
   • FGD Guidance: A test-time guidance mechanism that uses the gradient of the FGD metric to adjust the predicted mean during the denoising process. It is activated only during contact phases to prevent foot-skating while preserving jump motions.

3. Anti-Freezing Reward:

   • The Problem: Physics-based rewards naturally favor "freezing" or static motions because static poses have zero physical violations and are easy for the imitation policy to track.
   • The Solution: An additional reward term that penalizes low motion magnitude. It calculates the velocity and acceleration of pose and translation parameters, encouraging the model to generate dynamic, large-amplitude movements while maintaining physical validity.

3. Key Contributions

1. Bridging the Skeleton-to-Mesh Gap: The paper identifies and addresses the critical disconnect between skeletal generation and mesh visualization, introducing a framework that enforces mesh-level physical constraints directly during generation.
2. Physics-Based RLFT: The authors propose a novel method to distill physical priors from a non-differentiable physics simulator into a differentiable diffusion model using an imitation policy as a reward evaluator. This avoids the need for complex motion projection during inference.
3. Comprehensive Reward System: The design includes a tripartite reward system (Imitation, FGD, and Anti-Freezing) that balances physical realism with aesthetic dynamism, solving the "freezing motion" bias common in physics-guided generation.
4. Generalization: The method is model-agnostic and successfully fine-tunes multiple SOTA diffusion models (EDGE, POPDG, GENMO, Bailando++).

4. Experimental Results

The method was evaluated on the AIST++ and PopDanceSet datasets.

• Physical Plausibility:
  • Penetration Rate: Reduced by 49% compared to the baseline EDGE model.
  • Foot-Ground Contact (PFC): Significantly improved (lower scores indicate better contact), outperforming all baselines.
  • FGD: Drastically reduced foot-ground deviation, indicating realistic foot placement.
• Aesthetic Quality & Dynamics:
  • User Studies: Human evaluators rated the generated dances as having higher overall aesthetic quality and physical plausibility compared to baselines.
  • Motion Magnitude: The anti-freezing reward successfully prevented the model from collapsing into static poses, maintaining motion dynamics comparable to ground truth.
• Efficiency:
  • Unlike motion projection methods (e.g., PhysDiff) which require iterative simulation during inference (slowing down generation and causing failure rates up to 55% in complex scenarios), Skeleton2Stage requires no additional simulation during inference. The physical constraints are internalized by the model, resulting in a 100% success rate and significantly faster inference times (3s vs. 30s-120s for projection methods).

5. Significance and Impact

• Real-World Applicability: By ensuring physically plausible mesh visualizations, this method makes generated dances suitable for film, gaming, and VR applications where body collisions and foot-ground interactions must be realistic.
• Paradigm Shift: It moves the field from "generate then fix" (post-processing) to "generate correctly" (intrinsic physical constraints), offering a more robust and efficient solution.
• Addressing FID Limitations: The paper critically discusses the unreliability of FID (Fréchet Inception Distance) for dance generation, showing that lower FID can sometimes correlate with worse physical plausibility (e.g., jittery motions). The authors advocate for physics-based metrics (Penetration Rate, PFC) and user studies as more reliable evaluation standards.

In conclusion, Skeleton2Stage represents a significant advancement in dance generation by successfully integrating complex physical laws into generative AI, ensuring that the resulting animations are not only rhythmically aligned but also physically valid and visually realistic.