From teleop data to a deployed robot policy — in 30 days.
Managed Pi-0 fine-tuning for robotics teams that don't want to run JAX, openpi, sim eval, and inference serving themselves. We handle the four bits: record, train, eval, deploy. You ship a robot.
Already running. Already pushing weights.
Our reference implementation fine-tunes Pi-0 on a public LeRobot dataset and pushes the adapter to HuggingFace. Same pipeline runs on your data, on your robot, in your pilot.
Four bits to a deployed robot policy.
Most teams have one or two of these working. Almost no one has all four. Fourbit gives you all four — your robot's data goes in, a working deployed policy comes out.
Record
Teleop episodes in. LeRobot format, your robot.
Upload your LeRobot dataset, or use ours. We validate sync, joint coverage, action quality — and surface the bad episodes before training wastes GPU time.
Train
Pi-0, OpenVLA, ACT, Diffusion Policy.
Fine-tune the right policy for your robot. Bimanual SO-101, ALOHA, custom rigs. Curated configs, per-second H100 billing, automatic resume on preemption.
Eval
Sim success rate. Real-robot success rate. Both.
Side-by-side rollouts: base policy vs your tune. gym-aloha, gym-pusht, or your own sim. Plus on-hardware eval on a matching SO-101 in our lab.
Deploy
Policy endpoint your robot can hit. 30-day on-call.
Low-latency inference over WebRTC/gRPC. Docker image you can run locally. A fourbit engineer in your Slack for the first 30 days post-deploy.
Don't have hardware? We'll ship you a robot.
The Fourbit Pilot Kit lands on your desk: a bimanual SO-101 leader-follower setup, calibrated cameras, a starter teleop dataset, and your first fine-tune on us. Standardized hardware means a productized pipeline — your pilot runs the same code we ran for ours.
- Bimanual SO-101 setup, pre-calibrated, plug-and-play
- Two RGB cameras + mounting rig
- One free Pi-0 fine-tune + sim eval scorecard
- 30-day on-call engineer in your Slack
- $5K credit applied against your first paid month
- Bimanual SO-101 (2× arms + grippers)~14-DOF
- Cameras2× RGB, 224×224
- Starter dataset50 episodes, 1 task
- First fine-tunePi-0 LoRA
- Sim evalgym-aloha rollouts
- Real-robot evalOn your bimanual
- Deploy targetLocal Docker + private endpoint
- On-call engineer30 days
Hardware is yours to keep. Continue paying after the pilot for ongoing fine-tunes, evals, and deploys at retainer pricing.
Boring infrastructure. Interesting policies.
LeRobot-native
Datasets in LeRobot format land in our Modal volumes ready to train. Bimanual, single-arm, custom action spaces — we adapt the recipe, not your data.
Modal-backed compute
H100 / A100 capacity with per-second billing. Spot for non-urgent runs, on-demand for pilot deadlines. Resumes from checkpoint on preemption.
Eval harness, not vibes
gym-aloha and gym-pusht out of the box. We replicate your sim or hardware on our side and score your tune against the base model — same task, same seeds.
Real-robot deploy
Low-latency policy serving over WebRTC/gRPC. Runs as a Docker image on your hardware or as a private fourbit endpoint. <50ms control loop.
One pilot. 30 days. A working policy.
We onboard a small number of robotics teams per quarter. If you have a teleop dataset (or want one of our pilot kits), let's talk.