TacMan-Turbo: Proactive Tactile Control for Robust and Efficient Articulated Object Manipulation

1. Problem Statement

Robots operating in human-centric environments must manipulate articulated objects (e.g., drawers, doors, microwave ovens) that often have uncertain or unknown kinematic structures. Current approaches face a fundamental trade-off between effectiveness (robustness to structural uncertainty) and efficiency (speed and smoothness):

• Model-based methods: Rely on pre-defined kinematic models. They are efficient but brittle; they fail when the actual object structure deviates from the model.
• Reactive Tactile methods (e.g., Tac-Man): Use tactile feedback to compensate for contact deviations without prior models. They are highly robust but inefficient. They operate on an "execute-and-recover" cycle: the robot moves, detects an error, stops to correct the deviation, and then resumes. This results in slow, zig-zag trajectories and significant time overhead.

The Goal: Develop a control framework that achieves the robustness of tactile-informed methods while eliminating the efficiency loss caused by reactive stopping and recovery cycles.

2. Methodology: TacMan-Turbo

TacMan-Turbo introduces a proactive tactile control framework that unifies the execution and recovery phases into a continuous control loop. The core innovation is a temporal interpretation of tactile signals.

A. Core Insight

Instead of treating contact deviations merely as instantaneous spatial errors requiring immediate compensation, the system analyzes sequential tactile observations as time-series data. This allows the robot to extract local kinematic information (motion patterns) from the contact history, enabling it to predict future object states and adjust velocities proactively.

B. Technical Framework

The system operates through three interconnected components:

1. In-hand Pose Estimation:

   • Using a geometry-based tactile sensor (e.g., GelSight), the system tracks contact points between the gripper and the object handle.
   • It estimates the relative transformation between the gripper and the handle by minimizing the distance between current contact points and a reference state (using the Kabsch algorithm).

2. Handle Pose Prediction:

   • Leveraging the temporal sequence of estimated poses, the system predicts the future handle position.
   • It employs a Constant Velocity Model (assuming locally smooth trajectories) to extrapolate the next pose ($\tilde{T}_{h_{i+1}}$) based on the two most recent pose estimates.
   • This prediction allows the controller to anticipate where the handle will be, rather than just reacting to where it is.

3. Proactive Offset Velocity Computation:

   • The controller calculates an offset velocity ($u^i_g$) to add to a base velocity ($u^0_g$).
   • The objective is to minimize predicted future contact deviations, not just current ones.
   • The system solves an optimization problem to align the gripper's future trajectory with the predicted handle trajectory, subject to robot joint velocity limits.
   • This eliminates the need for dedicated "recovery" phases; corrections are integrated seamlessly into the forward motion.

3. Key Contributions

1. Temporal Interpretation of Tactile Signals: A novel approach that treats contact deviations as time-series data to infer local kinematics, enabling prediction without pre-provided models.
2. Proactive Control Framework: Reformulates articulated object manipulation from a reactive pose-correction problem to a proactive velocity-modulation problem. This unifies the execute-and-recover cycle into a single continuous process.
3. Training-Free and Model-Free: The method does not require reinforcement learning training, prior kinematic models, or manual parameter tuning (like PID gains). It relies on analytical geometric interpretation.
4. Robustness to Disturbances: The system can handle unexpected disturbances (e.g., external forces) by continuously adjusting without stopping, unlike reactive baselines.

4. Experimental Results

The authors evaluated TacMan-Turbo on 200 diverse simulated articulated objects (prismatic, revolute, and complex Bézier curve trajectories) and real-world experiments with physical robots (Kinova Gen3 arm with GelSight sensors).

• Success Rate: Achieved 100% success rate across all 200 simulated objects and real-world tests, matching the robustness of the previous state-of-the-art (Tac-Man).
• Time Efficiency:
  • Order-of-magnitude improvement: TacMan-Turbo completed tasks in ~10 seconds on average, compared to ~250 seconds for Tac-Man.
  • Statistical significance: $p < 0.0001$.
  • Speedup factors: ~15x for drawers, ~11x for microwaves, and ~18x for vises.
• Action Efficiency:
  • TacMan-Turbo maintained high action efficiency (motion directly contributing to task progress) by eliminating non-productive recovery stops.
  • Tac-Man suffered from low efficiency due to alternating execution and recovery phases.
• Trajectory Smoothness:
  • Measured by time-weighted joint jerk. TacMan-Turbo produced significantly smoother trajectories, avoiding the abrupt accelerations/decelerations characteristic of reactive recovery cycles.
  • Mean jerk was reduced from ~4291 $s^{-3}$ (Tac-Man) to ~110 $s^{-3}$ (TacMan-Turbo).
• Generalization: Successfully manipulated everyday household objects (battery docks, power switches, coffee makers, sewing machines) not seen during training or simulation.

5. Significance

• Resolving the Trade-off: TacMan-Turbo successfully breaks the long-standing trade-off between effectiveness and efficiency in articulated object manipulation. It proves that robots can be both robust to unknown structures and highly efficient.
• Practical Deployment: By removing the need for precise kinematic models and training data, this approach is a "drop-in" upgrade for robots operating in dynamic, unstructured human environments.
• Scalability: The method is compatible with various tactile sensing modalities (force-based or geometry-based) and can be adapted to higher sampling frequencies for even greater efficiency.
• Future Impact: This work paves the way for robots to perform complex household tasks (opening drawers, turning knobs, unscrewing caps) with human-like speed and reliability, a critical step toward general-purpose home robotics.