research
Our work centers on the safety of robot interaction — how robots can act in physical contact with people and objects without putting them at risk. We pursue it along four connected directions, from the mechanisms of human motion, through perception and control at the moment of contact, to the policies and the hardware that make safe interaction possible in the first place.
Human Motion Mechanisms 人体运动机理
Robots that work in contact with people need a model of what a human body can and cannot tolerate. Skin-level kinematics are not enough: the same external motion can correspond to very different internal loads depending on posture, muscle activation, and joint state.
We build biomechanically grounded models of human motion — musculoskeletal structure, not just skeletal keypoints — so that an interaction can be evaluated against what is actually happening inside the body. The open question we care about is how to recover these internal quantities from the sparse, noisy signals available outside a laboratory.
Representative work
Papers in this direction will be listed here as they appear.
Omni-modal Perception & Force Control 全模态感知与高频力控
Contact is where a robot’s senses disagree most. Vision is occluded exactly when touch becomes informative; force readings are fast but local; proprioception drifts. Each modality is partial, and they run at wildly different rates.
We work on fusing vision, touch, force, and proprioception into a single coherent estimate of contact state, and on closing control loops over it fast enough to matter — contact events unfold in milliseconds, which puts hard real-time requirements on both the perception and the control stack.
Representative work
Papers in this direction will be listed here as they appear.
Safe Interaction Policy 安全交互策略
A policy that is safe in expectation is not the same as a policy that is safe. Rare contact events dominate the risk, and they are precisely the events least represented in training data.
We study how to learn and verify interaction policies — across navigation, locomotion, and manipulation — whose safety properties hold under physical contact rather than only in free space, and that degrade gracefully when the world falls outside what they were trained on.
Representative work
Papers in this direction will be listed here as they appear.
Novel Embodiments & Sensors 新型本体与传感器
Some safety properties are cheaper to build into a body than to compute at runtime. Compliance, intrinsic force sensing, and a well-chosen morphology can rule out entire failure modes before any controller runs.
We design novel embodiments and sensors so that safe physical interaction becomes a property of the hardware, reducing how much the software has to guarantee on its own. This closes the loop with the other three directions: better bodies make perception and control tractable.
Representative work
Papers in this direction will be listed here as they appear.