Tianyou Zuo左天佑

I am developing a low-cost, extensible robotic arm designed to assist laboratory users with chemistry experiments, starting with titration.

Reading chemical reactions The research explores how sharp changes near reaction-stage endpoints can help a robot recognize the current stage and anticipate transitions. The aim is to guide experimental actions using both laboratory protocols and observed chemical phenomena.

Classroom deployment & a planned data cycle The arm has been deployed in laboratory classes at Peking University’s College of Chemistry and Molecular Engineering. I aim to use routine replicate titrations to collect data for stage-recognition and robot-action models, linking classroom experiments, model training, and further experimental evaluation.

Two coupled feedback loops · research goal The goal is to couple robot-execution feedback—whether an operation was carried out as intended—with chemical-state feedback—how the reaction is progressing. Together, these loops would let the arm adapt its operations to the experiment rather than simply replay a fixed sequence.

We study task-oriented active physical perception: using deliberate interactions to infer the object properties that matter for the manipulation task that follows.

From interaction to grasp selection The robot actively interacts with an object or its environment and infers the object’s physical properties from sensor feedback. These estimates guide the selection of grasp points suited to the intended manipulation, connecting what the robot senses to how it will use the object.

Experimental findings In our experiments, this approach produces more robust grasps and improves the success rate of subsequent manipulation.

We developed a neural force-inversion approach that estimates forces and contact locations from visible fingertip deformation, enriching compatible UMI demonstrations with physical interaction information.

Shared geometry, material-aware inference The model is trained using simulation and real force-calibration data for a shared UMI fingertip geometry, accounting for material differences including 90A and 95A TPU variants. Applying it requires compatibility with the fingertip material, print characteristics such as infill, and available visual observations.

Enriching data without re-recording Existing recordings can be augmented without re-recording their demonstrations. The added force and contact-location signals are model-derived estimates, providing additional physical information for learning from the original data.

Labware grasping: hardware tested, force feedback next We have tested compliant, enveloping grasping of laboratory glassware, including beakers, on hardware. Closed-loop grasp control using estimated force is in development, with the aim of combining physical compliance and force feedback for gentle handling of labware.