OhhO Train
Turn demonstrations into policies.
OhhO Data collects the demonstrations; OhhO Serve runs the model — Train is the engine in between. Fine-tune Vision-Language-Action, imitation-learning and reinforcement-learning policies from your own episodes or in simulation, track every run, and export a deployment-ready checkpoint.
- Fine-tune VLA, ACT, diffusion & RL policies
- Trains from OhhO Data or in simulation
- Built on the open OmniVLA engine
- Experiment tracking & sweeps (W&B)
- Exports to Serve + ONNX for Fleet OTA
What you get
A policy is only as good as the loop that produced it. OhhO Train is the standardized training engine for embodied AI — the productized OmniVLA engine — so you fine-tune real models without stitching together a different toolchain for every method.
Train covers the methods that matter: behavior cloning and VLA fine-tuning (SmolVLA, ACT, diffusion, OpenVLA) on the LeRobot datasets you record with OhhO Data, plus reinforcement learning in Isaac Lab with domain randomization for sim-to-real. A continual-learning loop can re-train as new episodes and tasks arrive, with replay, multi-objective rewards and AI-judged self-evaluation.
Every run is tracked — losses, success rate, evaluation — with Weights & Biases sweeps to find good hyperparameters. When a checkpoint passes verification, Train exports it in the format the rest of the stack expects: a checkpoint OhhO Serve loads directly, or an ONNX policy OhhO Fleet ships over the air.
Built to do the hard parts for you
Behavior cloning, VLA fine-tuning (SmolVLA / ACT / diffusion / OpenVLA), offline RL and on-policy RL — selected by config, not a rewrite.
Point Train at a LeRobot dataset from OhhO Data, or generate experience in Gazebo and Isaac Sim with domain randomization for sim-to-real.
A post-training loop re-trains as new episodes and tasks arrive, with prioritized replay and outcome-stratified episodic memory.
Multi-objective rewards (task, safety, efficiency, smoothness) plus vision rewards and AI judges score behavior, not just loss.
Losses, success rate and eval stream to Weights & Biases; Bayesian sweeps search hyperparameters for you.
Verified checkpoints export to OhhO Serve and to ONNX for OhhO Fleet OTA, with hardware-aware execution providers baked in.
How it works
Pick a policy type and point Train at an OhhO Data dataset — or a simulation task.
Launch the run; losses, success rate and eval stream to W&B in real time.
Best-of-N evaluation with hard safety and reachability checks gates what's allowed to ship.
Push the checkpoint to OhhO Serve, or export ONNX for OhhO Fleet to roll out.
Specifications
- Methods
- BC, VLA fine-tune, ACT, diffusion, offline + online RL
- Data sources
- OhhO Data (LeRobot), Gazebo, Isaac Sim
- Foundation
- Open-source OmniVLA engine
- Tracking
- Weights & Biases + Bayesian sweeps
- Hardware
- Your GPU — we help you size it or recommend a rig
- Export
- Serve checkpoint + ONNX for Fleet OTA
Which plan do I need for OhhO Train?
We recommend the Builder plan. Builder gives you cloud training with full experiment tracking — enough to fine-tune a first real policy. Teams running sweeps or standing up a continual-learning loop want Fleet; enterprises training on their own cluster choose Forge.
Common questions
Yes. Train is built on the open-source OmniVLA engine, so your training code and checkpoints are portable — you're never locked in.
No. You can train entirely in simulation with domain randomization, then fine-tune on real OhhO Data episodes for sim-to-real transfer.
Works better together
Ready to build with OhhO Train?
Start free and simulate first — no hardware required. Upgrade when you're ready to deploy.