Robotics Engineer — AI Manipulation, Perception & Control

About the Role

We build AI-first mixed-case palletizing, truck unloading and truck loading — robots that look at a stream of never-before-seen packages and stack them into pallets, cages and trailers, or unstack a loaded trailer onto a conveyor, reliably and fast, with no per-box programming.

You'll own this system end to end — the learned stack and the deterministic spine beneath it — from 3D perception through grasping, placement and motion planning to commissioning on real KUKA cells at the customer's site.

The site is a new automated warehouse we're building bottom-up with SF Express at their logistics center in Suzhou. These are our first deployments — production logistics lines, not pilots — and you'll ship them. The architecture is a deterministic spine with learned components upstream: everything that has to hold rate and never surprise the safety PLC stays explicit and debuggable. The stack runs on a Jetson AGX Orin real-time gateway plus an x86 NVIDIA GPU workstation, and success is measured in pick success rate, cycle time and load rate on a live line — not in papers.

You'll work with engineers from NASA, UC Berkeley, Stanford, CMU, Apple, and Google, and with KUKA on the robot side, based at the Suzhou site or our Shenzhen office. The role is open from early-career to senior, in vision, planning, learning or integration. What doesn't change with level is wanting to be in the cell, not only the lab: if you've taken an arm application through commissioning on a customer floor and want to be back there, you'll feel at home — and strong lab or self-built manipulation work counts too. This role may require travel to some of the world's most exciting factories — Midea, Mercedes-Benz, SF Express, and others — mostly in southern China.

What You'll Do

• Build the perception stack — structured-light 3D vision (overhead, eye-in-hand and place cameras), hand-eye calibration, class-agnostic segmentation, 6-DoF cuboid pose fits of arbitrary parcels.
• Design grasp selection — analytic suction scoring plus learned grasp generation — and the zoned vacuum end-effector logic for rigid, film-wrapped, and deformable packaging.
• Build the placement brain — online 3D bin packing into pallets, roll-cages and transport-stable trailer walls: stability gate, reach-in feasibility, warehouse-system (WMS) lookahead.
• Sequence trailer unloading — full-face wall scan, stability-safe unstack order, next-box vision masked behind arm motion, the carrier platform advancing row by row.
• Plan collision-free, time-optimal motion with cuRobo over a live signed-distance field (ESDF) built from cage and trailer scans, latency budgeted against arm motion.
• Drive KUKA arms at 250 Hz — joint targets to KR C5 (KSS 8.x / iiQKA.OS2) over RSI: AxisCorr, IPOC timing, Ruckig jerk-limited trajectories, watchdog, timeouts.
• Write KRL, submit-interpreter and EthernetKRL (EKI) programs in WorkVisual — EXT-mode start chain, RSI context, gripper I/O — and synchronize placement to the belt encoder.
• Integrate and commission the cell — safety PLC (ISO 10218 / 13849), light curtains, telescopic conveyor, carrier platform, KUKA.Sim, fake-parcel validation, run-in with KUKA and SF.
• Close the loop — record every pick, run the exception ladder (detect, auto-recover, operator takeover, route-away), feed the data flywheel and retrain the models.

What We're Looking For

• Strong Python and C++.
• Hands-on robotic manipulation experience — an industrial arm, a research lab, or systems you built yourself.
• 3D computer vision — point clouds, depth cameras, and pose estimation.
• Motion planning and control fundamentals, and machine learning with PyTorch (segmentation, grasping, or policy learning).
• Fluent with agentic coding tools like Claude Code as part of your daily workflow.
• Proficient English is required; professional Chinese is strongly preferred (our partners and deployment sites are in China).
• You like being hands-on with hardware and on the factory floor.
• Bonus: KUKA or other industrial-robot integration (RSI, KRL, EtherCAT), palletizing / depalletizing / bin-packing, or structured-light 3D cameras (Photoneo, Zivid, Mech-Mind, or KUKA vision).
• Bonus: NVIDIA Isaac Sim, ROS 2, cuRobo / OMPL, LeRobot, imitation learning / VLAs, or NVIDIA Jetson.
• Bonus: you've built multi-agent systems, or have deep experience with agentic coding tools like Claude Code.

Why Join Us

Work directly with the CTO and founding engineers, and own a real product going live with world-class industrial partners — KUKA, Midea, Mercedes-Benz, and SF Express. You'll get direct advising from leading faculty at UC Berkeley BAIR and from researchers at Physical Intelligence and other company partners. The team feels like a family: Friday-night events and occasional retreats — racing, beach days, and more (last summer we took the whole team to Hawaii!). There's freedom to explore, with room for open-ended work that aligns your skills with the direction of our research. Pay is set by experience within the band — performance-based salary plus bonus — with meaningful equity. We can assist eligible candidates with work-permit arrangements.