Bliss Reading Group - July 6

Join us as we dive deeper into our ongoing mini-track focusing on robot perception for physical interactions! This session highlights the vital shift from sensing contact to effectively acting upon it.

Featured Paper: DexForce: Extracting Force-Informed Actions from Kinesthetic Demonstrations for Dexterous Manipulation (Chen et al., 2025)

Imitation learning for dexterous manipulation often struggles with a key challenge: capturing the nuances of finger movements, particularly the force exerted during interactions. Traditional methods, such as teleoperation with motion retargeting, fail to address the importance of force data, which can have a significant impact on outcomes.

DexForce steps in to tackle this issue by incorporating direct guidance (kinesthetic teaching) to teach the robot hand while simultaneously recording contact forces from fingertip sensors. These measurements are crucial for generating "force-informed actions" during policy learning.

Results Highlight: Policies trained using force-informed actions boast an impressive 76% success rate across six manipulation tasks, contrasting sharply with the near-zero success rate of policies that ignore force information.

This presentation complements the insights from SonicBoom, showcasing how acoustic sensing can identify contact location. In contrast, DexForce emphasizes that understanding force application is pivotal for developing effective policies.

Discussion Points:

• Which tasks demonstrate heightened sensitivity to force information?
• Is it feasible to gather force-aware demonstrations without kinesthetic teaching?
• What revelations does the near-zero performance of current imitation learning pipelines unveil?

Come ready for engaging discussions and thought-provoking questions! 🎉