I gave a browser a button that says "train a robot," wired it to Nebius Serverless AI, and got back walking robots, rollout videos, and downloadable policies.
That is Sim2Policy, my entry for the Nebius Serverless Challenge 2026.
Repository: https://github.com/andygolubev/nebius-serverless-challenge-2026 (MIT).
Site: https://sim-policy-trainer-challenge.info (For now, it is accessible only to users with @nebius.com email addresses during the judging period.)
The idea
Training a robot policy in the cloud is easy. Keeping the proof after the machine vanishes is hard. Serverless GPU jobs spin up, train, and disappear — and usually take the evidence with them.
Sim2Policy fixes that. Each run trains a policy on a bounded Serverless AI Job and leaves behind a durable, reproducible result. You never touch a GPU, an image, or a config file — you just pick an example and watch it learn.
Every run leaves behind:
- A trained checkpoint — the policy, saved before the job exits.
- A rollout video — the robot going from wobbling to walking.
- Evaluation metrics — reward, episode length, runtime, and cost.
- A downloadable policy bundle — checkpoint, config, versions, and checksums in one file.
Under the hood there are two training engines behind the same simple button: a dependable CPU path for the classic control robots, and a GPU-parallel path for the big quadruped and humanoid. You never have to know which one ran — the platform picks it for you.
The architecture
A tiny always-on VM in the Nebius cloud runs the app and launches disposable training jobs. The user stays in the browser. The cloud does the rest.
Pick an example, press go
The hosted gallery offers seven ready-to-run robots — from classic MuJoCo controls on CPU to the Go1 quadruped and G1 humanoid on GPU. Click one and it trains live on Nebius.
Bring your own robot
You can also upload your own robot model, snap together a task and a scene, and train it — all inside guardrails, no code required.
The result
When a run finishes, you get a full result page: the lifecycle, the key numbers, a progression video of the robot learning to move, and a one-click policy bundle download.
And the headline number that made me smile: the flagship Go1 run trained end-to-end on a Nebius H100 — 102.4 million steps in about 10 minutes for roughly $0.52. The CPU examples cost about $0.02 each. This is real GPU reinforcement learning at coffee-money prices.
Behind the scenes it is all just Nebius jobs — you can watch them run in the console, on H100, L40S, and CPU shapes side by side:
You can see details:
You can check metrics:
and logs:
Under the hood
For the curious, here is the stack:
- Simulation: MuJoCo through Gymnasium, and MuJoCo Playground / MJX on JAX for GPU-parallel physics.
- Learning: PPO — Stable-Baselines3 on CPU for the classic robots, Brax PPO on GPU for the Go1 and G1.
- Hardware: NVIDIA H100 and L40S GPUs for the JAX path; CPU shapes for classic control. Each job picks the right one automatically.
- Platform: FastAPI + React app on a single-node k3s VM, deployed with ArgoCD, all provisioned by OpenTofu on Nebius.
How a run works: the job runs PPO for a fixed number of steps, checkpoints to S3 as it goes, then renders rollout videos and runs a deterministic evaluation before publishing the result. If a job is interrupted, the durable checkpoints let it resume instead of starting over.
What you can configure: pick the robot/example and an optional random seed. Bring your own robot and you also choose the locomotion task (stand-and-balance or walk-forward), a scene preset (flat arena, ramp, hurdles, steps), and a few scene objects. The backend, image, GPU, and command stay locked down — you get choices, not footguns.
Why it works
Everything runs on Nebius Serverless AI Jobs — you pay for the burst of compute and nothing else. The app is served from a single small VM, deployed by GitOps, with secrets kept out of Git. The heavy lifting is entirely serverless, entirely disposable, and entirely reproducible.
The README covers setup, hardware, expected outputs, and cost. Everything you see above is in the repo:
https://github.com/andygolubev/nebius-serverless-challenge-2026
Top comments (0)