In classical robotics, physical contact from a human is often treated as a "disturbance"—an external noise that a control system should resist to maintain its trajectory. However, a growing body of research suggests that these forces are not noise, but a rich, high-bandwidth communication channel. By interpreting physical human-robot interaction (pHRI) as a signal of intent, robots can learn to adapt their objectives, synchronize their movements, and collaborate more naturally with human partners.
From Disturbance Rejection to Information Extraction
The fundamental shift in this research area is the move away from viewing physical contact as something to be ignored or suppressed. Bajcsy et al. argue that physical pushes are actually corrections to a robot’s internal objective function. When a human nudges a robot arm, they are providing information about what the robot should be doing. By treating these interactions as active feedback, the authors demonstrate that robots can refine their understanding of a task in real-time.
This perspective transforms the robot from a rigid machine into a responsive collaborator. Losey et al. show that when robots interpret physical interactions as communication, they require less human effort and interaction time to complete tasks. Users in their studies perceived these "learning" robots as more understanding and easier to work with compared to traditional systems that merely resisted external forces.
Proxies for Human Intent
Inferring intent requires a way to "feel" the human. While dedicated force-torque sensors are common, researchers have developed several clever proxies to detect human guidance without specialized hardware:
- Velocity Tracking Errors: Shao et al. propose using the discrepancy between a robot's commanded motion and its actual motion as a proxy for human force. This "sensorless" approach avoids the noise and cost of physical force sensors while providing a clean signal for co-manipulating heavy objects.
- Kinematic Manipulability: Mohammadi et al. demonstrate that changes in a robot's posture—specifically its "manipulability ellipsoids"—can reveal a user's desired direction. By monitoring how external forces alter the robot's configuration, the system can infer walking directions or velocities.
- Tactile Deformation: For delicate tasks like handovers, Rayamane et al. and Naveen Kuppuswamy use vision-based tactile sensors. By tracking the displacement of markers or internal membrane patterns within a soft "skin," these robots can detect the subtle shear forces of a human pulling an object away, signaling the exact moment to release the grip.
Learning from Physical Corrections
Once a robot can sense interaction, it must decide how to update its behavior. This is often framed as Learning from Corrections (LfC). Wang et al. describe this as a paradigm where a supervisor provides kinesthetic nudges during task execution to update the robot's task objectives on the fly.
To make these updates robust, researchers are employing sophisticated modeling techniques. Liu et al. use causal convolutions and spectrograms to process force data, treating it as a primary input modality alongside vision. This allows the robot to detect sudden shifts in human intent—such as an abrupt change in direction—that visual sensors might miss due to occlusion or latency. Similarly, He et al. employ Bayesian intent inference to maintain safety while ensuring the robot remains responsive to unpredictable human interventions.
Applications in Coordination and Guidance
The ability to infer intent through touch is critical for tasks requiring tight synchronization:
- Co-manipulation: In legged robotics, Bethala et al. decoupled intent inference (turning forces into velocity) from locomotion. This allowed a humanoid robot to carry a load with a human partner using only haptic cues, achieving coordination levels comparable to two humans working together.
- Robotic Guidance: For assistive technologies, Cui et al. developed a system for blind and visually impaired users where a quadruped robot learns the relationship between leash tension and the user's walking style, enabling personalized gait synchronization.
- Rehabilitation: In clinical settings, Vianello et al. use exoskeletons to link therapists and patients. By amplifying the subtle physical cues of a patient, the system allows a therapist to "feel" the patient's underlying motor intent through the robotic hardware.
By treating physical touch as a meaningful dialogue rather than a disturbance, these frameworks enable robots to move beyond pre-programmed trajectories and toward true, intuitive partnership with humans.
Key Foundational Work
Learning Robot Objectives from Physical Human Interaction (2017) — contributed 2 concept elements to this direction.
Self-learning and adaptation in a sensorimotor framework (2016) — contributed 1 concept elements to this direction.