Robot policy training

Motion.
With proof.

Kam turns a robot behavior into a scoped training job.

Hopper Rover · recorded policy trajectory, photoreal render

The product

Your robot in.
A tested policy out.

  1. 01

    Send

    Your model and a task.

  2. 02

    Train

    Isaac Sim, 4,096 robots in parallel.

  3. 03

    Judge

    MuJoCo, a second physics engine.

  4. 04

    Deliver

    One zip: the network, ROS 2 package, recipe and report.

What's in the zip

Every package we have built uses this layout: Skydio X2, Unitree A1, Courier Hex.

policy/policy.onnx
The trained neural network in ONNX, with its weights and sha256. The README gives the exact observation and action layout and the control rate.
jetson-build-engine.sh
Jetson deployment. Builds the TensorRT engine on your Jetson and exits 0 only if it matches the ONNX output (parity under 1e-3).
ros2/pf_policy/
ROS 2 package. Builds with colcon: the policy node, launch files, and a MuJoCo sim plant so you can drive it before it touches your robot.
README.md
Integration guide. Topics, joint order, PD gains, command ranges, the commands to run it, and the limits we found.
training/
The training recipe. Final PPO checkpoint, the resolved environment and agent settings (rewards, randomization, network), the Isaac Lab task, the learning curve, and RETRAIN.md to fine-tune it yourself.
gate/ selftest/ report.html
The test report. Every check with its value and limit, plus the canary self-test that proves the gate can still fail.
robot/ lineage/
Provenance. The robot model as trained, and a hashed record of each stage (import, train, gate, deploy) with its cost.
  • GATE PASS

    Skydio X2

    60 episodes, 0 crashes

    Report
  • GATE PASS

    Unitree A1

    72 episodes, 0 falls

  • GATE PASS

    Hopper Rover

    60 of 60 courses, 0 crashes

The gate

It reached the goal.
We still pulled it.

This rover never lands. Our evaluator rejects it.

Side view of one test course.

startgoal,

Loading the run…

driving rotors on touchdown wall 3 cm speed bump gap the gate checks: green landed, red flown over One held-out course. Heights exaggerated.
Recorded simulation · v18a · MuJoCo

Current policy

Trained in Isaac Sim.
Judged in MuJoCo.

Gate courses
60/60
Crashes
0
Gaps flown
0

Simulation only. Hardware testing is next.

Raw from the simulator

Start

Your robot.
Our next challenge.

  1. Your model. MJCF or URDF with meshes.
  2. Its limits. Fill in the sheet or send the robot card.
  3. What it senses. State, control rate, computer.
  4. The behavior. In numbers.
Pilot terms and limits
  • You keep your robot model, your data, the trained policy and the TensorRT package.
  • We may use anonymized test results. Never your name or model without your written OK.
  • Provided as is, no warranty for hardware use. Either side can stop at any time.
  • Simulation-validated only; not tested on a physical robot.
  • No safety-critical use without your own safety sign-off, hardware testing and a person able to take over.

Plain-language draft; a lawyer has not reviewed it yet. We confirm the price in writing before any work starts.

Trained so far: quadrupeds, multirotors with one thrust actuator per rotor, and our own drive-and-fly rover. No passing humanoid, arm or wheeled-only results yet. Unsure? Send it anyway and pick "other".

01You
02Robot and model
Robot type
Model format optional

GitHub, Drive, Dropbox or any link we can download. Leave empty to send it by email.

Your MJCF or URDF with its meshes, zipped. It is stored privately; only our team and training pipeline can read it.

Skip anything the model already contains. We read masses and limits from the file and ask only for what is missing.

03What it can sense, and where it runs
State your robot can estimate tick all that apply

We train on this state, not on camera or lidar images.

04The behavior