Robotics Engineer

About Ubundi

Ubundi is a South African venture studio building human-centered AI. First Motive, one of our ventures, builds the data engine for Physical AI — the ground-truth infrastructure that turns real-world, multimodal capture into clean, replayable, train-ready datasets, and the robots that learn from them.

We already have a platform that captures data and a pipeline that turns it into trained policies. What we need now is the person who closes the loop — who takes a learned model and makes a robot act on it.

The Mission

Own the seam between model and robot. Take the VLA/VTLA policies our data engine produces and deploy them onto real and simulated robots — orchestrating tasks, serving inference in the loop, and making autonomous behaviour actually happen on hardware.

What You'll Do

• Build the task brain — orchestrate multi-step robot behaviour per task: sequencing, arbitration, recovery
• Serve learned VLA/VTLA policies into the robot in real time, closing the perception → action loop
• Bridge the model side (datasets, training) to the robot side (our ROS2 platform)
• Turn research policies into reliable, repeatable autonomous runs on sim and real hardware
• Define how a new task is specified, launched, and evaluated — so autonomy scales across verticals
• Debug the hard seam: timing, latency, safety, and failure recovery between policy and actuation

What You'll Bring

• Strong ROS2 — enough to plug into a production robot platform without hand-holding
• Real-time policy deployment and ML inference — serving models into a control loop
• Orchestration: behaviour trees, state machines, task planning and arbitration
• Python, and C++ where the loop demands it
• A foot in both camps — fluent enough in ML to deploy a model, in robotics to run it on hardware
• Systems thinking and failure-mode instinct — you reason about what breaks before it breaks

Nice To Have

• LeRobot, RLDS, or other embodied-AI / imitation-learning stacks
• Experience deploying VLA/VTLA or other robot-learning policies
• MoveIt, ros2_control, or motion-planning depth
• Sim-to-real transfer, domain randomization
• Foxglove, rosbag, or replay-driven evaluation workflows

What We Offer

• A genuine greenfield seam to own — the model→robot layer is yours to architect from inception
• A flat team of strong engineers; you own your lane, no one micromanages it
• "I am because we are" — high trust, radical candor, respect for life outside work
• Hybrid out of Stellenbosch, with studio access when the hardware needs you
• Work at the live frontier of Physical AI, on the bottleneck that matters