ML Cloud Engineer — Teleoperation & Inference

About the Role

Intuition is a Silicon Valley robotics startup backed by some of the most prestigious investors in tech, connected to Founders Fund, MiniMax, Raspberry Pi, and others. Our team has backgrounds from Stanford University, UC Berkeley BAIR, Fudan AI, CMU Robotics Institute, and Google, and we work with industrial partners including KUKA, UR, Midea, Mercedes Benz, P&G, IKEA, and SF Express Logistics.

You'll join the Rapid Prototyping team at our Shenzhen deployment center: robots from many vendors arrive with an application, we build the station platform around them, and the solution ships. Every station shares one platform: a height-adjustable torso, two arms at ALOHA spacing, a top-view neck camera, fish-eye wrist cameras, a mid-height torso camera, and one Jetson Orin. V0 is a stationary single-arm station; V1 adds the second arm, one cable to the wall, and wheels that lock solid; V2 goes mobile with wheeled navigation, battery, and 5G.

You'll own the three things that have to stay smooth at once: the teleoperation network, the robot data flow, and the cloud model-inference runtime. On the station, a Jetson Orin answers an industrial arm's UDP state stream at 300 Hz while a GELLO leader arm sends commands at 50 Hz. One Ethernet cable is the station's whole link to the internet. Everything you build lives on the far side of that cable, and the arm must never feel it.

When a policy fails on a station, a remote teleoperator takes over, corrects it, and hands back; the correction then goes into the next training run. That failure-recovery mode is DAgger, and your first hard problem is making it work over the internet: very low latency, and an arm that never loses its smoothness. In V2 the stations go mobile, and the cable to the internet becomes a 5G module.

Then you'll build the system underneath all of it: teleoperators over the internet, driving robots that move as if the teleoperator were standing beside them. The profile: you have chased a latency bug across a network, a kernel, and a GPU, and want the next one to end at a robot arm.

What You'll Do

• Sustain a smooth teleoperation network from a remote teleoperator to the robot: low latency, and no jitter the arm can feel.
• Own the robot data flow: four camera streams and arm state leave each Jetson Orin on one Ethernet cable, ready for training.
• Run the cloud model-inference runtime: GPU serving that returns the robot's next action before the control loop needs it.
• Implement the DAgger failure-recovery mode over the internet at very low latency: a teleoperator corrects a failing policy; the arm stays smooth.
• Ensure the cloud model stays available at the local site, and own how that holds when the link degrades.
• Build the distributed system that lets teleoperators control robots over the internet, as smoothly as with the teleoperator in the room.
• Design around the robot-side loop: a 50 Hz GELLO leader arm, a 300 Hz arm, and a failsafe tripped by late replies.
• Instrument every station: latency, jitter, dropped cycles, and inference round-trips, visible before a teleoperator notices them.
• Keep cloud inference and control reliable when V2 swaps the wall Ethernet for a 5G module.

What We're Looking For

• No hard prerequisites to apply: we hire on what you have built and kept running, not on titles.
• Distributed systems you have built and run under load: message ordering, backpressure, clock drift, and what happens when a link drops.
• Low-latency networking down to the packet: UDP, jitter, loss, and buffering, and what each does to a control loop.
• Strong Python on Linux: you profile, pin a process to a core, and measure latency end to end before believing a number.
• GPU inference serving: you have put a model behind an endpoint and kept its tail latency where the caller needs it.
• Exposure to robotics or real-time systems: you know why a 300 Hz loop cannot wait for a slow reply.
• Based in Shenzhen or ready to relocate.
• Bonus: you have driven a robot over a network, or built a teleoperation link of any kind.
• Bonus: you know DAgger and imitation learning well enough to design the recovery loop, not just the transport.
• Bonus: agentic coding tools like Claude Code or Codex are already part of your workflow.

Why Join Us

When a station ships from the deployment center, the teleoperator's hands reach its arm through your network, and the model reaches it through your runtime. That stays true from one stationary arm to mobile robots on 5G. You'll join a small new team where the network and inference side is yours, from the packet leaving the Jetson Orin to the model served from the cloud. Interns earn uncapped project bonuses, and top-performing interns may receive full-time offers and the opportunity for stock options. Full-time pay is set individually, with equity case by case and accommodation included. Direct mentorship comes from UC Berkeley BAIR professors and researchers at Physical Intelligence. Some interns are encouraged to pursue open-ended exploration, letting their own strengths shape what they work on. If you want low-latency distributed systems with a robot arm at the end of the wire, this is it.