Drives, control loops and fail-closed safety, from the bus to the robot.
Five years at SS Innovations. On the SSi Mantra surgical arm: bench-level motion control and the EtherCAT master, which I wrote. Alongside it, I lead the technical side of a four-person programme on a Unitree G1 humanoid: coordinator, safety path, navigation bridge, voice. 2,000+ commits across its last two platform generations.
Five years on SSi Mantra, India's first indigenous surgical robot, a 12-DOF system. Motion control on its arm, validated at bench level: cascaded current/velocity/position loops with dual-encoder feedback, gain scheduling and friction compensation on frameless-torque-motor and harmonic-reducer joints, working directly with compliance, lost motion and two-mass resonance, for smooth low-speed surgical motion. ELMO and Ingenia servo controllers, Bode/Nichols on the bench for 7+ joints, and torque sensing on individual actuators.
A pySOEM EtherCAT master for the arm: PDO mapping, the CiA 402 (DS402) drive state machine, bring-up and tuning across 7+ joints, and a bench commissioning flow that cut drive bring-up time by more than half. I wrote the master. Every byte on the bus accounted for.
Harnessing layouts, calibration procedures, repeatability checks and written acceptance tests (step-wise pass/fail, sign-off) in a regulated medical-device environment. That habit carried over to the humanoid, where the safety path now surfaces a failure to the operator instead of passing it silently.
A personal EtherCAT bench of four CiA 402 drives with a torque-ceiling clamp on the command path, now at jog stage. The rule on it: a drive value counts as known only when my declaration and the drive's own readback agree, and nothing reaches an actuator until it is bench-verified.
On the humanoid the walking policy owns balance, so the safety work is about catching faults and refusing motion rather than braking. Both paths below run on the real robot.
Safety is a separate 100Hz thread pinned to a core, requesting SCHED_FIFO, watching thermal and battery thresholds. It never auto-damps, by design: the walking policy controls balance, so a damp from a side thread mid-walk causes the fall it exists to prevent. If the safety thread dies, the coordinator releases the arms, flags safety_ok=false, raises the alert and leaves the damp to the operator. I rebuilt that watchdog path so a failure surfaces to the operator instead of passing silently.
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 about 10cm from it, eyes on, gantry-suspended. On the nav chain, my part 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: a localization kill put the bridge into emergency halt in a measured 1.53s. No state feed, no drive: it publishes only 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 Unitree G1 with Inspire dexterous hands, 41 degrees of freedom, on a four-person programme where I lead the technical side: tech stack, hardware accessories and software. 2,000+ commits across the last two of five platform generations. Monolithic proof-of-concept to a split-compute dual-Jetson platform, with a behavior-tree coordinator, a safety thread that requests the coordinator's highest SCHED_FIFO priority, and a voice pipeline that runs on the robot. Coordinator, safety, voice and the nav bridge all run on the real robot.
Priority arbitration Emergency > Teleop > Voice > Idle in a
BehaviorTree.CPP v4 coordinator I designed and built. When teleop ends it ramps the arms off rt/arm_sdk at 50Hz;
rt/lowcmd belongs to the walking policy. Hand control publishes at 100Hz; hand-state feedback and vision results
cross processes over a lock-free seqlock shared-memory interface, memory_order_acquire and all, and world state
goes over ZeroMQ.
Whisper large-v3-turbo → an LLM (today a Gemma-4 26B MoE, with a custom fine-tuned 14B as rollback) → Kokoro TTS. The pipeline was originally squeezed onto a 16GB Orin NX (three models taking turns on one GPU behind a shared lock), and now runs on an AGX Orin 64GB with room to breathe. STT, TTS and VAD candidates each went through my own head-to-head shootout harness. A custom speculative-decoding path (draft model, batched verify, KV realignment) was built from scratch and benchmarked, then retired behind a flag when it lost on the number. In a design A/B, tag-routing beat LLM tool-calling by a measured 690ms per turn. Separately, I built a C++ speaker-ID pipeline that knows who is talking: Silero VAD → TitaNet embeddings, with temporal evidence gating.
A DDS bridge streaming joint targets at a commanded 500Hz (2ms timer), with cubic-Hermite interpolation, fed by MoveIt planners across OMPL, Pilz and CHOMP. Velocity scaling had no effect on execution in that pipeline, so I applied it by retiming the planned trajectory's timestamps in the bridge: a measured 10× execution-duration range from the same planned path.
Voice latency. Before/after numbers committed to git on an earlier generation of the voice stack. The wins came from KV-cache reuse, sentence-level LLM/TTS interleaving, and a GPU lock that hands each model the whole GPU on its turn, not bigger hardware. On today's stack, the same measure runs p50 2.82s, p95 4.82s over 40 live turns.
At NUS in Dr. Marcelo Ang's lab, I designed and fabricated soft pneumatic actuators whose asymmetric bending is baked into the material itself: two elastomers cured as a viscosity gradient. Springer IAS 2022. An IEEE hexapod paper before that.
Hiring for drives, motion control or robot safety? Write to me. I am open to relocation.