Overview
This research introduces CONTACT (CONtact-aware TACTile Learning), a framework investigating the utility of tactile sensing in robotic disassembly operations. The study systematically evaluates different sensing modalities, specifically focusing on the performance of a vision-only approach compared to approaches augmented with two distinct tactile sensing configurations: tactile RGB (TacRGB) and tactile force field (TacFF).
Research Context
Robotic disassembly is characterized by contact-rich interactions. Successful manipulation within these operations is dependent not only on geometric alignment but also on force-dependent state transitions. While vision-based robotics often perform effectively in structured environments, their reliability can diminish in scenarios involving tight tolerances, dominant contact, or deformable materials. The research aimed to determine the specific role tactile sensing plays in these challenging aspects of robotic disassembly.
Approach
The study employed a unified learning framework to compare three sensing configurations: Vision Only, Vision + tactile RGB (TacRGB), and Vision + tactile force field (TacFF). This comparison was conducted across both simulated and real-world experimental settings.
Simulation Experiments
For the simulation phase, five distinct rigid-body disassembly tasks were constructed. These tasks were designed with progressively increasing levels of geometric constraints and extraction difficulty to thoroughly test the robustness of the different sensing approaches.
Real-World Experiments
The real-world evaluation involved the design and execution of five tasks. Among these, three tasks focused on rigid scenarios, while two tasks specifically addressed deformable scenarios. These real-world tasks were chosen to evaluate manipulation processes that are highly dependent on contact interactions.
Findings
- Across both the simulation and real-world experimental setups, policies incorporating TacFF consistently demonstrated the highest success rates.
- The performance gains obtained with TacFF were particularly prominent in contexts characterized by contact-dependent interactions and in scenarios involving deformable objects.
- A comparison between TacRGB and TacFF revealed that TacFF-based policies outperformed those relying on TacRGB.
- The direct concatenation, or 'naive fusion,' of TacRGB and TacFF inputs resulted in policies that underperformed when compared to policies utilizing either TacRGB or TacFF individually. This indicated that simply combining these modalities could potentially dilute task-relevant force information rather than enhancing it.
- The results suggest that tactile sensing plays a critical, task-dependent role in the success of robotic disassembly operations.
- Specifically, structured force-field representations, as implemented through TacFF, were found to be particularly effective in contact-dominated environments.