Hi, I am Yi Gao (高熠), a third-year undergraduate in the SWUFE-UD Institute of Data Science, a dual-degree program between the University of Delaware and Southwestern University of Finance and Economics. I am pursuing a B.S. in Information Systems, with a background in data science. I am based in Chengdu, China, and expect to graduate in 2028.

I am interested in robot learning for physical interaction: how robots can respond to contact while retaining the motion and support needed to complete a task. I currently work with Pengxiang Ding at SymBiosis Robotics, where I independently lead SoftSONIC, a project on learning compliant adaptation of a frozen humanoid motion controller. Previously, I developed context-grounded robot agents for industrial patrol at Lenovo Robotics Research Institute (Shanghai).

I am seeking a remote research collaboration in contact-rich manipulation or humanoid control, available immediately for 30 hours/week. I can contribute to policy implementation, reproducible simulation experiments, data pipelines, and deployment-oriented evaluation.

CV (PDF) · Email · GitHub · Reading notes (Xiaohongshu)

Research

SoftSONIC: Learning Compliant Adaptation of a Frozen Humanoid Motion Prior

SymBiosis Robotics · Research Intern & Ongoing Remote Collaboration, May 2026 – Present
Research guidance: Pengxiang Ding

Motion tracking specifies what a robot should do, but physical contact can require it to deviate from that motion. Can we learn a reusable contact response on top of an existing whole-body controller, without retraining its motion prior? SoftSONIC investigates this question by separating the nominal motion skill from a lightweight learned adaptation.

• Method: freeze the released SONIC encoder and decoder, and learn a bounded 64-dimensional latent residual from proprioceptive and previous-action history. PPO uses SoftMimic-inspired compliant motion targets and SONIC’s native motion-quality rewards. Deployment requires no explicit external-force measurement for the residual actor.
• Shared-policy study: trained and evaluated a shared residual policy on 32 training motions in simulation, with paired force/no-force comparisons and separate tests on references excluded from residual-policy training. Controlled studies of residual action spaces, observation conditioning, and reward variants examine the tradeoff between compliant response and nominal motion quality.
• Real-robot progress: deployed the adapter through a 50 Hz C++/TensorRT pipeline on a Unitree G1. Hardware trials show qualitative yielding under manual pulls during standing, walking, and selected motion replays, including a backward-walking reference excluded from the residual-policy training set. These observations remain preliminary, and response strength varies across motions.
• My contribution: independently led the research question, method and data-pipeline development, experiment design and analysis, and deployment validation, with guidance from senior collaborators.

Current work scales compliance augmentation to thousands of loco-manipulation motion sources and tests broader shared-policy generalization. Large-scale training is in progress. Task-level evaluation with a fixed GR00T policy is a next step; the current hardware observations are not a calibrated stiffness or task-success benchmark.

Updated October 8, 2026. Research in progress.

Real-robot demonstrations

Manual interaction during G1 motion replay. The videos show standing, walking, and a door-opening motion from the residual training set, plus a backward-walking reference excluded from residual-policy training that also exhibits yielding. This is a preliminary qualitative observation; broader generalization evaluation is ongoing.

Recorded with an overhead safety harness and manual perturbations. Training labels refer only to the residual policy’s training set, not to the frozen SONIC models’ pretraining data. These clips demonstrate observed behavior, rather than a measured stiffness or task-success benchmark.

Experience

Context-Grounded Robot Agent for Industrial Patrol

Lenovo Robotics Research Institute (Shanghai), Solution & Development Center
Research Intern, July – December 2025

• Developed a robot agent prototype that combines visual observations with location/time-specific rules and action protocols to support traceable industrial-patrol decisions.
• Collected and organized 200+ real-world video clips using a DEEPRobotics X30 quadruped; built structured annotations and a VLM evaluation workflow comparing QwenVL and GLM-family models.
• Designed hierarchical context retrieval and drafted CE4Patrol, a manuscript on multi-layer context reasoning for industrial anomaly inspection.

VLA Inference for Retail Mobile Manipulation

Technical Project, March – April 2026

Reproduced a RoboBenchMart VLA inference pipeline in ManiSkill, including asset preparation, scene generation, model serving, and evaluation-client integration for retail pick-and-place and open/close tasks.

Manuscripts

CE4Patrol: Multi-Layer Context Reasoning for Industrial Anomaly Inspection
Yi Gao, et al. · Manuscript