NemoClaw agents running Nemotron on Nebius Token Factory train the MicroDuck in NVIDIA Isaac Lab, judge it, tune it, verify it and stage it for the real robot — from one sentence in a chat window or one command in a terminal.
Why it matters
Getting a walking policy from simulation onto a small robot normally means launching training, watching reward curves, guessing hyperparameters, exporting, converting, checking interfaces and hoping it doesn't fall off the table. Here an agent does the loop, and you approve the one step that moves real motors.
Students, hackathon teams and robotics newcomers say what they want; the agent picks the task, settings and tools.
Policies are gated on fall rate, not reward alone. A failing policy resumes with gentler learning automatically.
Nemotron 3.5 Super VL looks at frames from the rollout video and critiques how the duck actually walks.
9 sim-to-real checks on every policy, and the robot is never loaded until you say the duck is on its stand.
Status replies are rendered from the real pipeline state and relayed verbatim — the agent can't invent progress.
Your GPU only simulates. All the reasoning runs in the cloud on Token Factory, so no GPU memory goes to the LLM.
Nebius Token Factory
Picking the task, reading reward curves, tuning, judging the gait from video and running the sim-to-real pipeline — through one OpenAI-compatible endpoint and one API key.
nvidia/Nemotron-3-Ultra-550b-a55b
Reads the failed run — fall rate, episode length, reward trend — and picks the retry's learning rate and entropy, with a one-line reason. Answers in about 2 seconds; bounded, with a fixed-rule fallback.
nvidia/nemotron-3-super-120b-a12b
Plans and calls tools for the Claw and the Duck pilot. Reliable multi-step tool calling, 262K context.
nvidia/Nemotron-3.5-Super-VL
Watches six frames of the rollout video and says what's wrong with the gait — answers in about 3 seconds.
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B
Five times cheaper for quick turns — switch the Claw to it live from the console's model switcher.
Paste it in the console: checked against /v1/models, stored at mode 600, pushed into the sandboxes, never shown again.
Agents run in NemoClaw / OpenShell sandboxes with deny-by-default egress — allowed to reach Token Factory and the Isaac control server.
GPU work stays on your own machine; a full train → verify run costs only the handful of agent turns that drive it.
How it works
The Duck pilot — a Hermes agent in its own NemoClaw sandbox — starts the pipeline on the host, reports on it, and stops at the one step a human must approve.
$ ./duck ship train ✓ ─ gate · ─ play ✓ ─ export ✓ ─ verify ✓ ─ deploy ✓ MicroDuck · flat — Done (dry run) Policy 2026-10-29_16-00-25 · iteration 750 (best existing) Verify ✅ 9/9 checks · 786 KB · 61→14 · Δ onnx/torch 3e-06 Bundle isaac_sim/_deploy/IsaacContrib-Velocity-Flat-MicroDuck_20261029-183351 Robot not configured (dry run)
Architecture
The green paths are inference calls to Token Factory. Tool calls reach the Isaac control server through the OpenShell L7 proxy; the real robot is only loaded after a human confirms.
Click to open the full-size diagram.
Results so far
MicroDuck on flat ground, trained by the pipeline and exported for the robot.
Quick start
Needs an NVIDIA GPU, Isaac Lab with the MicroDuck task, NemoClaw, and a Token Factory API key.
echo "NEBIUS_API_KEY=<your key>" >> ~/.openclaw/.env chmod 600 ~/.openclaw/.env ./demo.sh check # key, model, GPU, ports ./demo.sh up # → http://localhost:8800 ./demo.sh down
./duck run -n 1500 --tries 2 # full pipeline ./duck ship # best checkpoint → robot ./duck watch # live status card ./duck deploy # asks: on its stand? ./duck chat # talk to the Duck pilot