Active Extrinsic Contact Sensing: Application to General Peg-in-Hole Insertion (2110.03555v2)

Abstract: We propose a method that actively estimates contact location between a grasped rigid object and its environment and uses this as input to a peg-in-hole insertion policy. An estimation model and an active tactile feedback controller work collaboratively to estimate the external contacts accurately. The controller helps the estimation model get a better estimate by regulating a consistent contact mode. The better estimation makes it easier for the controller to regulate the contact. We then train an object-agnostic insertion policy that learns to use the series of contact estimates to guide the insertion of an unseen peg into a hole. In contrast with previous works that learn a policy directly from tactile signals, since this policy is in contact configuration space, it can be learned directly in simulation. Lastly, we demonstrate and evaluate the active extrinsic contact line estimation and the trained insertion policy together in a real experiment. We show that the proposed method inserts various-shaped test objects with higher success rates and fewer insertion attempts than previous work with end-to-end approaches. See supplementary video and results at https://sites.google.com/view/active-extrinsic-contact.

Active Extrinsic Contact Sensing and its Application to Peg-in-Hole Insertion

The paper, "Active Extrinsic Contact Sensing: Application to General Peg-in-Hole Insertion," introduces an innovative approach to address tactile sensing and dexterous manipulation tasks using active extrinsic contact sensing. This method accurately estimates the contact location between a grasped rigid object and its environment, serving as a pivotal input to the peg-in-hole insertion policy. Researchers Sangwoon Kim and Alberto Rodriguez propose a collaborative system comprising an estimation model and an active tactile feedback controller that work synergistically to refine the estimation of external contacts.

The tactile feedback controller is designed to regulate a consistent contact mode, enhancing the estimation model's efficacy. By maintaining a stable environment for contact estimation, this controller facilitates more precise contact regulation, thereby improving the insertion policy's accuracy. Differing from traditional methods that interpret policies directly from raw tactile signals, this paper's strategy operates in the contact configuration space, allowing for simpler and computationally efficient simulation training.

Numerical Results

The results obtained from real-world experiments are noteworthy. The proposed method demonstrated increased success rates and reduced insertion attempts across various-shaped test objects compared to earlier direct sensory approaches. Specifically, the approach achieved higher than 95% success rates in challenging scenarios, showing significant improvement in efficiency and reliability.

Key Components and Contributions

1. Factor Graph-Based Estimation: Utilizing incremental smoothing and mapping (iSAM), the paper integrates robot proprioception with tactile measurements to estimate contact lines accurately. The factor graph processes both gripper-object relative displacement and proprioceptive data, enabling real-time estimation updates.
2. Active Tactile Feedback Controller: This component strategically guides the object around regulated contact, creating conditions conducive to accurate estimation. Through a sequence of controlled movements—push-down, rocking, and pivoting—the controller achieves stable extrinsic contact modes.
3. Reinforcement Learning (RL) Policy: Trained in a low-dimensional simulation environment, this policy utilizes contact-line estimates as inputs. The ability to simulate the contact configuration space simplifies the training process, enhancing generalization to unseen objects.
4. Experimentation and Validation: The framework's validity is tested with a variety of peg shapes and holes, showcasing robust performance against initial misalignments.

Practical and Theoretical Implications

Practically, the successful implementation of active extrinsic contact sensing in peg-in-hole tasks paves the way for integrating these tactile manipulation techniques into robotic assembly lines and manufacturing processes where precision and adaptability are paramount. Theoretically, the development of a cooperative estimation-controller framework that interacts and adjusts dynamically marks an advancement in the fields of tactile sensing and robot learning.

Future Prospects

The paper suggests avenues for expanding the framework to more complex manipulation tasks beyond flat surfaces, such as point contacts or non-stationary contacts. Moreover, exploring applications in unstructured environments could achieve generalization in robotic operations, heralding advancements in autonomous systems and adaptive robotic intelligence. The reinforcement learning strategy could benefit further from integrating multi-modal sensory inputs that enhance robustness and adaptability in diverse scenarios.

In summary, this paper contributes significant advancements in understanding and implementing tactile sensing frameworks for robotic manipulation tasks. Through robust experimental validation, the potential for further developments in AI-driven robotic systems is illuminated, promising transformative steps in tactile robotics and automated precision engineering.