DPTG: diffusion policy with tactile feasibility guidance.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Fusion

Background:

• Integrating vision and tactile sensing is crucial for robotic manipulation in contact-rich environments.
• Current fusion methods often misrepresent tactile data as continuous motion, ignoring its role in providing physical constraints and phase cues.

Purpose of the Study:

• To propose a novel framework, DPTG, that reformulates tactile sensing as a physical feasibility constraint.
• To improve the safety and stability of robotic interactions by better utilizing tactile information.

Main Methods:

• Developed the DPTG framework where actions are sampled from a vision-driven diffusion policy and guided by a tactile feasibility classifier.
• Introduced an adaptive guidance schedule derived from feasibility scores to selectively apply tactile constraints.
• Trained the feasibility classifier as a standalone module for potential reuse across tasks.

Main Results:

• DPTG demonstrated improved success rates in robotic manipulation tasks.
• The framework successfully reduced peak forces during contact interactions.
• Experiments in simulation and real-world scenarios confirmed enhanced safety and stability compared to baseline methods.

Conclusions:

• The DPTG framework offers a more effective approach to integrating tactile sensing in robotics.
• Reformulating tactile sensing as a constraint leads to safer, more stable, and efficient robotic manipulation.
• The standalone tactile feasibility classifier enables broader applicability and reusability.