A Unitree G1 arrives as hardware and an SDK. The layer that turns it into a system is mine: the coordinator, the safety path, the bridge that drives it, voice, behaviour. Torn down and rebuilt five times. Four years of surgical-robot motion control came first. That's it in the video.
A Unitree G1 with Inspire dexterous hands, 41 degrees of freedom, on a four-person programme where the software is mine: 1,841 commits, single author, torn down and rebuilt five times. Monolithic proof-of-concept to split-compute dual-Jetson platform, with a behavior-tree coordinator, a safety thread that outranks everything, and a voice pipeline that never leaves the robot. All of it on hardware, none of it sim-only.
Priority arbitration Emergency > Teleop > Voice > Autonomous > Idle in a
BehaviorTree.CPP v4 coordinator. Joint commands stream at the 500Hz rate the DDS bus requires;
hand control publishes at 100Hz; processes exchange state over a hand-rolled lock-free seqlock shared-memory
interface, memory_order_acquire and all. Safety is its own 100Hz
SCHED_FIFO thread on a dedicated core: IMU pitch/roll envelopes, thermal and battery thresholds,
heartbeat watchdog that drops the robot to damped mode on a 3-miss.
Whisper large-v3-turbo INT8 → a custom fine-tuned 14B LLM → Kokoro TTS. The pipeline was originally squeezed onto a 16GB Orin NX (three models sharing one GPU via lock scheduling), and now runs on an AGX Orin 64GB with room to breathe. Every model earned its slot in a head-to-head shootout harness. The 14B runs custom speculative decoding (draft model, batched verify, KV realignment) written from scratch. Tag-routing beat LLM tool-calling by a measured 690ms per turn. Barge-in at 200ms. It also knows who is speaking: a C++ speaker-ID pipeline (Silero VAD → TitaNet embeddings, temporal evidence gating).
The chain came up end to end on hardware: a goal set 1.5m out and 90° off heading, the robot aligned, walked 1.10m, and stopped within 10cm of it, eyes on, gantry-suspended. My half is the bridge from planner to motors and everything guarding it, which is where the month actually went. Velocity and yaw-rate shaping under caps. Fail-closed on stale localization, measured at 1.53s from kill to emergency stop. No state feed publishes nothing but zeros. Armed only by an explicit flag, disarmed by default, walked up a rung ladder where every step needed a human to say go. The planner itself (A*, elastic band, pure pursuit) is an existing stack I drive, not one I wrote.
A 500Hz DDS bridge (2ms timer) with cubic-Hermite interpolation feeding 15 motion planners. MoveIt's velocity scaling was silently a no-op, traced to the source, then fixed by rescaling trajectory timestamps directly: a measured 10× execution-duration range from the same planned path.
Git-committed before/after numbers from the voice stack. The wins came from KV-cache reuse, sentence-level LLM/TTS interleaving, and GPU-lock scheduling, not bigger hardware.
Four years on SSi Mantra, India's first surgical robot, a 12-DOF cart. Motion control on its arm: cascaded PID with feed-forward, gain scheduling across the workload envelope, friction compensation for surgical-grade smoothness at sub-millimeter scales. Hands deep in harmonic drives, ELMO and Ingenia servo controllers, Bode/Nichols on the bench for 7+ joints.
A pySOEM-based EtherCAT stack: DS402/CiA 402 state machines, drive commissioning, characterisation routines for a friction-compensation rig. I wrote the master. Every byte on the bus accounted for.
Harnessing layouts, calibration procedures, repeatability and system-level safety compliance, the unglamorous work that lets a robot pass medical scrutiny. That paranoia is now the humanoid's safety architecture.
At NUS with Dr. Marcelo Ang's lab: asymmetric bending baked into the material itself: two elastomers cured as a viscosity gradient. Springer IAS 2022. An IEEE hexapod paper before that.
If any of this resonated (the humanoid, the bus, the bends without bones), write to me.