GOAT RACER ONE
An accessible Sim2Real reinforcement-learning platform — train a driving policy in NVIDIA IsaacLab, export ONNX, and deploy to a Jetson-class RC car in minutes. The same loop we bring to client engagements.
Why we built it
Sim2Real research has been gated behind expensive lab-grade robots and bespoke software stacks. We wanted to prove that a complete autonomy loop — simulator, training, transfer, deployment — could fit on hardware anyone can buy off the shelf.
GOAT RACER ONE is that proof. Higher on-board AI compute, real-time sensing, and a deliberately accessible chassis. The result is a platform where new policies move from notebook to physical run in minutes instead of weeks.
The platform
Off-the-shelf parts, deliberate choices, no exotic hardware.
Chassis & drive
Traxxas Ford Fiesta ST Rally 1/10 AWD platform. Sensored brushless motor, VESC controller with onboard IMU for smooth low-RPM control, digital steering servo.
Compute
NVIDIA Jetson Orin Nano 8 GB as the primary on-robot brain. Enough headroom for real-time inference plus auxiliary perception workloads.
Sensing
Intel RealSense D435 RGB-D camera for vision and depth. AprilTags handle waypoint identification. Field-of-view and USB throughput tuned for real-time use.
Power
Dual 2S 5000 mAh LiPo packs with low-voltage cut-off. RF kill switch for guaranteed remote shutoff during testing.
On-robot stack
Lightweight Python runtime over ZeroMQ + Protocol Buffers. No heavy ROS 2 dependency — fewer moving parts, easier to reason about.
Digital twin
URDF/USD assets, sensor models, and physics tuned to match the physical car. Fidelity here paid off downstream more than anywhere else in the stack.
What the loop looks like
Every engagement moves through the same five-stage pipeline. We own all of it.
1 · Model the world
URDF / USD assets, sensor models, physics parameters. Digital twins from Gaussian splats or CAD.
2 · Train the policy
IsaacLab + PPO. Domain randomization, reward shaping, and curriculum across thousands of parallel envs.
3 · Export & optimize
PyTorch → ONNX → TensorRT. INT8 / FP16 calibration. Profile every stage of the graph.
4 · Deploy to hardware
Jetson Orin or DGX Spark edge. Custom Python runtime over ZeroMQ + Protobuf. Integrated with VESC, RealSense, and custom sensors.
5 · Measure & iterate
Telemetry, replay, regression. Close the sim↔real gap with real-world data fed back into training.
Done · Hand off
Documentation, training for your team, and a reproducible pipeline so the loop keeps turning without us.
Digital twin fidelity proved essential. Correcting a single steering-model error in simulation produced significantly more stable real-world behavior than weeks of on-track tuning.
What we shipped
Policies trained entirely in simulation transferred cleanly enough to complete multiple waypoint runs on the physical platform. The build has been demonstrated live on NVIDIA Omniverse Livestreams, and the full pipeline — assets, training scripts, on-robot runtime — is reproducible end to end.
Where it's headed
Ongoing work: more robust localization, richer observation inputs, tighter camera characterization, and higher-speed autonomous driving. The platform doubles as our internal proving ground — every improvement we land here feeds directly into client engagements.
Want this loop running for your robot?
If you can describe the problem, we can scope the loop. URDF to Jetson — we own all of it.