DexForce: A Novel Approach to Dexterous Manipulation through Force-Informed Actions

Researchers from Stanford University have introduced DexForce, a novel approach to collecting dexterous manipulation demonstrations that leverages contact forces measured during kinesthetic demonstrations to compute force-informed actions for policy learning. This groundbreaking research aims to address the challenges of collecting high-quality demonstrations for contact-rich tasks, which are essential for advanced robotics applications. By incorporating force data into policy observations, DexForce achieves an impressive average success rate of 76% across six tasks, outperforming policies trained on actions that do not account for contact forces.

Key Takeaways:

• DexForce is a novel approach to collecting dexterous manipulation demonstrations that leverages contact forces measured during kinesthetic demonstrations.
• The research aims to address the challenges of collecting high-quality demonstrations for contact-rich tasks, which are essential for advanced robotics applications.
• DexForce incorporates force data into policy observations, which achieves an impressive average success rate of 76% across six tasks.
• Policies trained on actions that do not account for contact forces have near-zero success rates in contact-rich tasks.
• The research finds that using force data never hurts policy performance but helps most for tasks that require advanced levels of precision and coordination.
• The DexForce approach has significant implications for robotics and automation, enabling the development of more advanced robots that can perform complex tasks.
• The research has been peer-reviewed and published in the IEEE Robotics and Automation Letters journal.
• Funders for the research include Toyota Research Institute, Amazon, and Kwanjeong Fellowship.

Statistics:

• 76% average success rate of DexForce policies across six tasks.
• 6 tasks demonstrated in the study, including opening an AirPods case and unscrewing a nut.
• Near-zero success rates for policies trained on actions that do not account for contact forces.
• The research found that using force data helped most for tasks that require precision and coordination (76.2% success rate vs. 44% without force data).

Sources:

• Dexforce: Extracting Force-informed Actions From Kinesthetic Demonstrations for Dexterous Manipulation. Ieee Robotics and Automation Letters, 2025;10(6):6416-6423. Ieee Robotics and Automation Letters can be contacted at: Ieee-inst Electrical Electronics Engineers Inc, 445 Hoes Lane, Piscataway, NJ 08855-4141, USA.
• NewsRx. Investigators from Stanford University Report New Data on Robotics and Automation (Dexforce: Extracting Force-informed Actions From Kinesthetic Demonstrations for Dexterous Manipulation). Robotics & Machine Learning. June 16, 2025; p 176.