Robotics research · systems engineering

Chong Li

Incoming M.S. Research student in Mechanical Engineering at Carnegie Mellon University.

SJTU B.S. Electrical & Computer Engineering · 2026Michigan B.S.E. Mechanical Engineering · 2026CMU M.S. Mechanical Engineering, Research Track · 2026–2028

My training in ECE and mechanical engineering has led me to work across sensing, real-time software, dynamics, control, and learning. I focus on the difficult middle of robot learning: locating why a system fails, correcting the interface or architecture responsible, and validating the change on measurable timing, motion, data-integrity, and hardware criteria.

Contact CV (PDF)GitHub

Overview

I work where learned policies meet physical failure.

My strongest contributions have not been routine operation or data collection. They have come from debugging across layers: tracing degraded control to encoder contention, correcting frame transformations, separating slow policy inference from fast balance control, bounding discontinuous actions, and designing tools that expose corrupted data before it reaches training or hardware. I am interested in research where those engineering decisions are part of the scientific contribution.

Selected work

Failure analysis, original system work, and evidence.

Each project is organized around the technical problem I encountered, how I traced it, what I designed or changed, and the evidence available at that validation layer.

Humanoid systems · Undergraduate thesis

Making a GR00T–SONIC humanoid stack run in real time

I diagnosed and redesigned the interfaces that let an asynchronous VLA policy command a 50 Hz whole-body controller without allowing recording and perception workloads to destabilize control.

PICO 4 Ultra teleoperationIsaac-GR00T N1.7 PolicyServermulti-rate motion bridgeSONIC WBCMuJoCo / Unitree G1

Problems diagnosed and systems built

Root-cause diagnosis

Recording, not control, was starving the real-time loop.

When synchronized recording degraded control, I instrumented resource use, loop timing, and LowState age to isolate the bottleneck. The evidence traced the failure to CPU software H.264 contention rather than SONIC or network transport.

Real-time redesign

I removed the bottleneck without breaking the data contract.

I migrated video to the Jetson hardware encoder while preserving the LeRobot schema, then added 50 Hz scheduling, event latching, process-priority isolation, thread limits, and reduced frame copying. The sustained loop reached 49.9 Hz with 10.32 ms mean and 27.59 ms P99 LowState age.

Kinematic correction

I fixed wrist targets at the coordinate-system boundary.

I implemented source-aware VR-3PT wrist calibration, corrected the spatial rotation composition, and added per-arm fallback behavior so inconsistent VR input could not silently propagate into motion targets.

System architecture

I helped separate semantic latency from balance control.

I co-designed and implemented the GR00T action-to-SONIC motion-target bridge, exposing testable module boundaries and isolating 208.15 ms mean Orin policy inference from the nominal 50 Hz whole-body loop.

Imitation learning systems · PAIR research internship

Turning a brittle dual-Franka pipeline into a controlled experiment system

I traced recurring setup, action-continuity, and dataset failures that made physical policy rollouts unsafe or difficult to reproduce, then designed a reliability layer around dual-arm collection and ACT/Diffusion deployment.

dual GELLO controllersthree RealSense camerasROS 2 / ZMQLeRobot datasetsACT / Diffusion Policydual Franka FR3

Problems diagnosed and systems built

Startup discontinuity

I made policy activation begin from a verified state.

A learned policy could start from a pose inconsistent with its training episodes. I built dataset-derived start-target selection and a feedback-checked alignment state machine that uses conservative quintic motion, joint-specific tolerances, settle windows, replanning, and timeouts before execution is enabled.

Unsafe action chunks

I converted discontinuous predictions into bounded motion.

I designed a state-aware quintic time-stretch limiter against FR3 position, velocity, and acceleration envelopes, integrated it into both ACT and Diffusion executors, and made limiting the default. The executor records when and how much each action had to be stretched.

Hidden data faults

I built a non-destructive audit-and-repair system for LeRobot data.

The tool detects structural, video, numeric, synchronization, and robot-motion faults; keeps the raw dataset immutable; generates a reviewable repair plan; rejects unresolved unsafe episodes and implausibly large label-only corrections; and re-audits every derived dataset.

Fragile orchestration

I replaced manual launch sequences with supervised experiments.

I built dependency-ordered supervisors that check ports and installed configs, wait for fresh messages rather than topic names, sequence gripper homing to avoid startup timeouts, monitor every process, gate live deployment behind dry-run checks, and return the system to standby during reverse-order shutdown.

3D sensing · Manufacturing research

Measurement reliability for industrial 3D scanning

An experimental workflow for studying how geometry, surface condition, and scanner configuration affect defect visibility and metrology quality.

printed reference artifactsZivid 2+ MR60 / eviXscan Quadro+point-cloud registrationplane fittinggeometric error analysis

Problems diagnosed and systems built

Measurement problem

I turned scanner failure modes into controlled variables.

I designed 60 mm reference artifacts that isolate convex, concave, missing-corner, black, and white features, then built an acquisition workflow that separates geometry and surface-condition effects.

Analysis design

I replaced visual judgment with geometry-aware evidence.

I compared scanner presets and reflection-removal settings using fitted-plane RMSE, dimensional scaling, volumetric difference, and defect visibility rather than relying on a single aggregate accuracy number.

Reproducibility

I made the experiment repeatable across scanners and operators.

I documented the capture, registration, and analysis procedure for an accepted MSEC 2026 poster; the broader work received the Best Reproducibility Award at the U-M Data Science & AI Summit 2025.

Engineering approach

Research that another person can inspect and extend.

Studying electrical and computer engineering at SJTU and mechanical engineering at Michigan trained me to move between hardware, controls, sensing, and learning. It also taught me to translate different definitions of success into shared technical criteria.

Define reliability concretely

Turn broad goals into observable criteria: timing, action continuity, dataset integrity, repeatability, and explicit failure conditions.

Keep decisions traceable

Version experiments, isolate variables, preserve reversible changes, and document interfaces so another researcher can reproduce the system.

Validate at the right layer

Separate offline tests, simulation results, subsystem checks, and full hardware rollouts instead of allowing success at one layer to imply another.

Background

Mechanical systems, computation, and experimentation.

My training spans ECE and mechanical engineering rather than a single robotics layer. That breadth is most useful when a learning algorithm must be integrated with sensing, timing, communication, calibration, and physical constraints.

Education

2026–2028 expected

Carnegie Mellon University

M.S. Mechanical Engineering · Research Track

Incoming Fall 2026 · Pittsburgh, Pennsylvania

2024–2026

University of Michigan

B.S.E. Mechanical Engineering

GPA 3.819/4.0 · University Honors · Dean's List

2022–2026

Shanghai Jiao Tong University

B.S. Electrical and Computer Engineering

UM–SJTU Joint Institute · Graduated 2026

Additional engineering work

Wireless rover and embedded autonomy

Integrated RPLIDAR acquisition, Wi-Fi UDP transport, encoder interrupts, odometry, and motor control for the Mapleseed mobile-robot platform.

Experimental controls and vehicle dynamics

Designed and validated lead-lag control for a flexible-shaft apparatus and contributed MATLAB/CarSim chassis-dynamics analysis to the University of Michigan Solar Car team.

Core tools

Python, C/C++, PyTorch, ROS 2, MuJoCo, LeRobot, ZMQ, RealSense, point-cloud processing, embedded firmware, experimental design, and technical documentation.

Contact

Research collaboration at CMU

I am looking to contribute to a CMU group working on embodied AI, robot learning, humanoid or manipulation systems, and learned motion control. I can contribute both to policy experiments and to the hardware/software infrastructure required to run them reliably.

chongl2@andrew.cmu.edu GitHub ↗Download CV