OhhO Train

Turn demonstrations into policies.

• Category: Intelligence
• Accent: cyan
• Live app: Open training console

The training engine behind the stack. Fine-tune VLA, imitation and reinforcement-learning policies from your OhhO Data episodes or in simulation — then export straight to Serve and Fleet.

Overview (hero)

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.

Highlights

• RL policy ONNX export script in the public repository
• Tiny ACT CPU training on the OHH-86 sim loop (once a run is green)
• Prototype training console — simulated run UI
• Vision: VLA fine-tune, diffusion, SmolVLA, OpenVLA
• Vision: continual learning, W&B sweeps, Fleet ONNX 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. Because the physical arm is a position-controlled 6-DOF servo arm with a parallel jaw gripper (no force sensors or tactile skin), policies target position-controlled pick-and-place. Real evaluation benchmarks are tracked In progress (OHH-101) without fabricated success tables.
• Weights & Biases sweeps are not built. An ONNX file for Fleet over-the-air updates is not built. Fleet over-the-air updates are not built. Force limiting, tactile feedback, and in-hand manipulation are Vision: not built.

Features

• Many methods, one engine — Behavior cloning, VLA fine-tuning (SmolVLA / ACT / diffusion / OpenVLA), offline RL and on-policy RL — selected by config, not a rewrite.
• Trains from your data — Point Train at a LeRobot dataset from OhhO Data, or generate experience in Gazebo and Isaac Sim with domain randomization for sim-to-real.
• Position-controlled manipulation — Trained policies command joint-position targets and parallel jaw states for pick-and-place. The physical robot has no force or tactile feedback; force-limited contact and in-hand dexterity are Vision: not built.
• Continual learning — A post-training loop re-trains as new episodes and tasks arrive, with prioritized replay and outcome-stratified episodic memory.
• Rewards & self-evaluation — Multi-objective rewards (task, safety, efficiency, smoothness) plus vision rewards and AI judges score behavior, not just loss.
• Vision: not built — Weights & Biases sweeps are not built. Exporting ONNX for Fleet over-the-air updates is not built.

How it works

1. Choose method & data — Pick a policy type and point Train at an OhhO Data dataset — or a simulation task.
2. Tracking sweeps are not built — Weights & Biases sweeps are not part of this console.
3. Verify — Best-of-N evaluation with hard safety and reachability checks gates what's allowed to ship.
4. Fleet export is not built — There is no ONNX package for Fleet over-the-air updates.

Specs

Plans

There is no paid plan. The console demo is free. Hosted cloud is a waitlist, not a checkout.

FAQ

Is the training engine open? Yes. Train is built on the open-source OmniVLA engine, so your training code and checkpoints are portable — you're never locked in.

Do I need real-robot data to start? No. You can train entirely in simulation with domain randomization, then fine-tune on real OhhO Data episodes for sim-to-real transfer.

Does Train support force-controlled or tactile policies? No. The physical robot uses a position-controlled 6-DOF servo arm with a parallel jaw gripper and no force or tactile sensors. Policies output joint positions for pick-and-place. Force limiting, tactile feedback, and in-hand manipulation are Vision: not built.

Related products

• OhhO Data
• OhhO Serve
• OhhO Proof
• OhhO Market