MaskAdapt: Learning Flexible Motion Adaptation via Mask-Invariant Prior for Physics-Based Characters

1. Problem Statement

Creating interactive, physically realistic humanoid characters that can perform a wide range of motions remains a significant challenge in robotics and animation. While recent physics-based control methods have successfully extended pretrained controllers using residual learning (adding a secondary network to modify behavior without retraining the base), existing approaches face two critical limitations:

1. Instability under State Distribution Shifts: Base policies are typically trained on full observations. When a residual policy later modifies specific body parts, the resulting state distribution shifts significantly. The frozen base controller, never exposed to these "missing" or altered observations, often produces unstable or incoherent actions.
2. Limited Flexibility in Adaptation: Current composition methods are often restricted to fixed body regions or lack an integrated mechanism for semantic control (e.g., text-driven guidance). They struggle to dynamically switch adaptation targets within a single sequence or combine multiple distinct motions seamlessly.

2. Methodology: MaskAdapt

The authors propose MaskAdapt, a two-stage residual learning framework designed to create a robust motion prior and enable flexible, targeted adaptation.

Stage 1: Learning a Mask-Invariant Motion Prior (Base Policy)

The core innovation is training a base policy ($\pi$) that remains stable even when parts of the state observation are missing.

• Stochastic Observation Masking: During training, the system randomly masks specific body-part groups (e.g., left arm, legs) in the state input $s_t$. The masked state $\bar{s}_t$ is created by zeroing out the corresponding dimensions.
• Mask-Invariant (MI) Regularization: To prevent the policy from collapsing into repetitive or unstable behaviors when inputs are missing, a regularization term is introduced. The MI Loss minimizes the Kullback–Leibler (KL) divergence between the action distribution under full observation ($\pi(\cdot|s, m^0)$) and the action distribution under masked observation ($\pi(\cdot|\bar{s}, m)$).
  • Objective: $L = L_{PPO} + \lambda_{MI} L_{MI}$
  • Result: The base policy learns to generate consistent, diverse, and physically plausible motions regardless of which body parts are "hidden" from its perception. This creates a robust prior that anticipates future adaptation.

Stage 2: Residual Policy for Flexible Adaptation

A residual policy ($\psi$) is trained on top of the frozen base policy to modify only the targeted body parts.

• Dynamic Motion Composition: The residual policy receives a mask $m_t$ as a conditioning signal. It learns to generate residual actions ($\Delta a_t$) specifically for the masked regions while leaving the unmasked regions to follow the base policy. This allows for dynamic switching of adaptation targets (e.g., adapting legs first, then arms) within a single sequence.
• Text-Driven Partial Motion Tracking: For semantic control, the framework integrates a pre-trained autoregressive diffusion model (from DART).
  • Text commands are converted into kinematic trajectories for specific body parts.
  • The residual policy tracks these kinematic goals for the designated masked regions while the base policy maintains the underlying global motion (e.g., walking).
• Conditional Discriminator: A discriminator $D_\psi$ is conditioned on the mask $m$ to enforce motion realism specifically for the adapted regions, allowing for selective enforcement of physical plausibility.

3. Key Contributions

1. Mask-Invariant Training Paradigm: The introduction of stochastic masking and MI regularization to train a base policy that is inherently robust to partial observations, solving the stability issues common in residual learning.
2. Flexible Motion Composition: A method enabling dynamic, on-demand adaptation of different body parts within a single motion sequence, surpassing the fixed-region limitations of prior work (like CML).
3. Text-Driven Physical Control: A novel integration of text-conditioned kinematic generators with physics-based residual control, allowing for semantically guided partial motion tracking (e.g., "wave arms while walking").
4. Unified Framework: A single architecture that handles both dynamic motion composition and text-driven goal tracking, demonstrating superior adaptability and robustness.

4. Experimental Results

The method was evaluated on a simulated humanoid (22 joints, 63 DoFs) using datasets like LAFAN1, AMASS, and BABEL.

• Base Policy Robustness:
  • Metric: Normalized Shannon Entropy ($H_{norm}$) of frame visitation.
  • Result: The MaskAdapt base policy with MI loss achieved an entropy of 0.9025, nearly matching the unmasked baseline (AMP, 0.9124). In contrast, a base policy trained without MI loss dropped to 0.7448, indicating a collapse into limited motion patterns when parts were masked.
• Motion Composition:
  • Compared against Composite Motion Learning (CML), MaskAdapt achieved higher entropy scores in both adapted and preserved regions across three test cases (Jump+Kick, Locomotion+Rotate Arms, Aim+Sneak).
  • Qualitative results showed smoother transitions and better coordination than CML, which often exhibited delayed activation or artifacts.
• Partial Goal Tracking:
  • In text-driven tasks, MaskAdapt achieved lower tracking errors (0.055m) and pose errors (0.047m) compared to CML and "from-scratch" models.
  • It successfully maintained stable locomotion while performing complex arm motions (e.g., waving, punching) triggered by text, whereas baselines often lost gait stability.
• Real-World Validation: The framework was successfully deployed on a Unitree G1 real humanoid robot, adapting a walking motion with a punching behavior.

5. Significance

MaskAdapt represents a significant advancement in physics-based character control by decoupling the stability of the base motion from the flexibility of the adaptation.

• Robustness: By explicitly training the base policy to be invariant to missing observations, the system avoids the "catastrophic forgetting" and instability that plague traditional residual learning when the state space changes.
• Interactivity: The ability to switch adaptation targets dynamically and control them via natural language opens new avenues for interactive animation and robotic control, where characters can respond to complex, multi-part instructions in real-time.
• Efficiency: The approach allows for efficient reuse of pretrained skills without full retraining, making it scalable for generating diverse, high-quality physical behaviors.

In summary, MaskAdapt provides a robust, flexible, and semantically controllable framework for physics-based characters, bridging the gap between high-level semantic intent and low-level physical execution.