Work
Four flagships on real hardware, the research around them, and an index of everything else with the state it is genuinely in, including the work that was archived or came back null.
Assistive autonomy · Unitree GO2
Say “come here” and a physical GO2 turns toward the voice, finds the caller, walks over and sits, all onboard its Jetson Orin NX with no internet at runtime.
Live on the physical GO2
Sim-to-real locomotion · Isaac Lab
Learned GO2 locomotion carried from Isaac Lab to the physical robot through a gated pipeline. The first deploy exposed a failure simulation missed; in October 2026 a newer walking policy held the real GO2 through every stand gate after a scripted stand-up, then walked forward and made a partial turn on a free floor. It walks short of the command because of joint friction the simulator did not model.
4.77e-06 max ONNX vs PyTorch drift on the stand-v3-h25 export, against a 1e-4 gate
Holds a stand and walks on the physical GO2; friction-aware retrains in first hardware tests
Runtime reliability · ROS 2
Runtime monitoring, diagnosis and audited recovery commands inside a live ROS 2 graph on the GO2, with BlackBoxRS replaying the evidence offline afterwards.
1.16 ms mean detection latency through the ROS 2 callback path, in a single-process benchmark
Decision path ran on the live GO2; physical actuation is the next hardware gate
Bare-metal C · Arm Cortex-M4F
A real-time runtime in C written from the reset vector up, with the board's ESP32-S3 network chip isolated behind a budgeted link.
11 → 0 forced network-chip crashes, controller resets, in one 10,003,440-period Phase 2 soak
v1.0, validated on hardware
Selected research & systems
C++20 · CUDA · TensorRT
A Python person-perception pipeline rebuilt in C++20, CUDA and TensorRT for the GO2's Jetson Orin NX: hardware decode, pose, re-identification and tracking.
135 → 33.5 ms p95 latency under matched replay of recorded camera frames on the robot's Jetson Orin NX; it keeps the 30 fps rate on 12% of one CPU core instead of 108%
Explore project Private repository
Conversational GO2 · local LLM + VLM
A GO2 you talk to: on “Fetch, describe the room” it scans three views and describes each through its own speaker, and it holds a conversation on local models with no API key, while a safety envelope alone moves the robot.
22.4 to 31.2 s voice-triggered three-view room scans in armed runs on the physical GO2, none with a degraded subsystem
Explore project Demo Private repository
Autonomous driving · pre-registered study
In a log-replay planning harness, a trajectory predictor made the planner brake for parked cars. A pre-registered stopped-agent fix cut those brakes on fresh scenes; a dose-matched sham did not.
44.9% fewer unnecessary hard brakes on 8,140 fresh scenes (95% CI 37.7 to 51.4%)
Production model teardown
A parity-controlled teardown of openpilot's shipped driving model: on simulator-rendered input it goes blind while its exported uncertainty stays flat.
Flight recorder for ROS 2
An incident recorder that builds a reproducible evidence bundle and replays recorded robot data through the same detectors offline.
Demos
Establish what the simulator is actually telling you before believing a number that came out of it.
Find the moment a run became unrecoverable, and watch a policy's own internals for the signature of a controller outside its competence.
The layer between a working idea and a robot that does it in a room.
Simulation is where a policy is allowed to be right; the robot is where it has to be. The hardware work on this page runs on one physical GO2.
I build autonomous robots, make them run efficiently on real compute, and instrument them so failures can be understood and contained. The research page makes that argument in full.