Technical Summary: ConTrack

Problem Statement

Transferring human demonstrations to dexterous robot manipulation is hindered by a "kinematic gap" arising from morphological and actuation differences between human hands and robot end-effectors. While existing methods can retarget high-level intent or coarse motion, they struggle with long-horizon, contact-rich sequences where precise contact timing and object trajectories are critical.

The core challenge lies in the trade-off between task success (keeping objects on target trajectories) and motion fidelity (preserving demonstrated joint motion and contact patterns). Existing approaches often rely on hand-crafted reward functions with fixed weights. These fixed mixtures are brittle: they require per-sequence tuning, fail when interaction budgets are limited, and cannot adapt when a reference motion is physically unrealizable for the robot embodiment.

Methodology

The authors propose ConTrack, a Reinforcement Learning (RL) framework that treats object tracking as a constraint rather than a weighted reward component. This allows the system to allocate remaining control authority to motion fidelity while strictly adhering to task requirements. The framework consists of three primary components:

1. Online Adaptive Task–Style Mixing (Dual-Variable Control)

Instead of fixed reward weights, ConTrack formulates the problem as a Constrained Markov Decision Process (CMDP). It maximizes style fidelity ($J_s$) subject to a minimum task success threshold ($J_g \ge \alpha J_g^\star$).

• Mechanism: A scalar dual variable $\lambda$ is updated online based on the ratio of the current running average of task return ($\hat{J}_g$) to a running maximum estimate ($J_g^\star$).
• Adaptation: If the task performance drops below the target ratio $\alpha$, $\lambda$ increases, shifting optimization pressure toward task tracking. If performance exceeds the threshold, $\lambda$ decreases, allowing the policy to prioritize style fidelity. This dynamic adjustment occurs without manual per-sequence tuning.

2. Adaptive Mid-Trajectory Reset Library

Long-horizon tracking is inefficient when training always starts from the initial frame, as early failures dominate the rollout distribution.

• Policy-Reachable States: Unlike standard mid-trajectory resets that copy reference states (which can be physically inconsistent in contact-rich scenarios), ConTrack maintains a library of simulator states collected from policy-reachable rollouts.
• Dynamic Sampling: The library is refreshed online. When a rollout achieves a longer continuation length from a specific frame $k$ than previously recorded, the state at $k$ is updated.
• Difficulty-Aware Sampling: Reset frames are sampled based on a "survival ratio" (the average length of successful continuations from that frame relative to the remaining horizon). This concentrates training on the specific segments where the policy currently fails, effectively moving the training focus backward as the policy improves.

3. Contact Priors

To anchor the style objective, the framework incorporates reference contact annotations as auxiliary priors.

• Style Reward: The style reward includes terms for contact event overlap (matching binary contact events) and contact point distance (minimizing the distance between simulated and annotated contact points on specific hand links).
• Role: These priors ensure that even when the robot deviates from the reference joint trajectory to satisfy physics, it maintains the demonstrated interaction patterns (e.g., which links touch the object).

Key Contributions

1. Online Adaptive Mixing: A dual-variable update mechanism that automatically balances task tracking and motion style, eliminating the need for manual reward tuning across diverse clips.
2. Adaptive Reset Library: A method for stabilizing long-horizon learning by reusing policy-reachable simulator states, avoiding the physical inconsistencies of resetting to raw reference data.
3. Constrained Framework: A unified approach that transforms physically infeasible reference motions into executable robotic trajectories, validated across three benchmarks of increasing contact complexity.

Experimental Results

The authors evaluated ConTrack on three datasets: GRAB (bimanual rigid-object interaction), ARCTIC (bimanual articulated objects), and DexterHand (single-hand in-hand rotation).

• Performance: Under a fixed training budget (5,000 PPO updates), ConTrack significantly outperformed baselines (ManipTrans, DexMachina, SPIDER) in success rate (measured as progress to the end of the clip) and object pose accuracy.
• Fidelity: While maintaining high task success, ConTrack preserved joint and contact fidelity better than methods that prioritized task success alone.
• Ablations:
  • Removing the adaptive mixing (using fixed weights) resulted in lower progress and lower contact fidelity, confirming the necessity of dynamic trade-off control.
  • Removing the adaptive reset library (using fixed start-frame resets) significantly reduced progress, highlighting the importance of focusing learning on failure boundaries.
  • Removing contact priors degraded contact accuracy, demonstrating their role in resolving contact dynamics.
• Real-World Validation: The authors demonstrated the feasibility of the learned trajectories on a real-world bimanual xArm7 + xHand platform, successfully executing tasks such as tool usage and object rotation.

Significance and Claims

The paper claims that ConTrack addresses a critical bottleneck in dexterous manipulation: the inability of current systems to robustly transfer human demonstrations to robots when contact schedules are complex and kinematic gaps exist.

• Scalability: By treating task success as a constraint and using online adaptation, the framework scales with diverse tracking data without requiring per-sequence reward engineering.
• Robustness: The adaptive reset library enables learning long-horizon tasks that would otherwise be intractable due to the "curse of horizon" in contact-rich environments.
• Practicality: The system successfully bridges the sim-to-real gap, producing trajectories that are executable on real hardware while maintaining the "style" of human demonstration.

The authors note limitations, including the reliance on the quality of contact annotations and the fact that the constrained formulation uses a running maximum for normalization, which provides stable scaling but does not strictly guarantee constraint satisfaction in all theoretical edge cases. However, the empirical results suggest the approach is highly effective for practical deployment.