Robot learning, as a service
Who trains the policy when you have no robot-learning team?
We take one task on one of your robots from about 100 demos to a measured policy, in ten weeks. You get a sim twin, a trained controller with a safety filter, a fine-tuned open VLA on your own hardware, and an evaluation report. You keep all of it.
It starts with a two-week paid audit. Or write to shashank@hedwigai.com.
Is this for you?
You build one kind of robot for one job: food assembly, farm work, inspection, a logistics arm or a heavy machine. You are between seed and Series B. You want a learned policy on the robot, and you cannot hire the engineers who build one.
The models are no longer the hard part. Isaac Lab, MuJoCo, π0.5, GR00T and SmolVLA are all open. The hard part is making them work on your robot, and proving that they do.
What you get.
A calibrated sim twin of your robot
Your robot is loaded into MuJoCo and Isaac Lab. Its actuators are fitted to your motor logs. The same robot is checked in both engines, so a result does not depend on one simulator.
A low-level controller with a safety filter
The controller is trained with reinforcement learning in the twin. It exposes a small skill API with stated rates. A safety layer enforces joint limits and the e-stop between the policy and the motors.
A fine-tuned open VLA on your own hardware
We test π0.5, GR00T and SmolVLA on your demos and keep the one that does best. The chosen model is fine-tuned and exported to Jetson Thor, or to the GPU you already run.
An evaluation report you can trust
Every success rate comes with a confidence interval. Policies are compared with sequential A/B tests on the real robot. The report says how many trials each number rests on.
Add-on: a world model as a policy evaluator. Once you have 20 to 60 hours of robot video, we post-train a world model to rank candidate policies before they reach the robot. We check its ranking against paired real runs first. We use it to rank policies, never as a simulator.
How it works.
- Audit2 weeks
- We learn your robot, task, cameras, compute and data. We check your demos and their licences, agree the success metric and a safe test protocol, and measure your current success rate. You get a written go or no-go and a fixed scope. The audit fee is credited against the sprint.
- Policy sprint10 weeks
- One task on one robot, from 100 to 200 demos to a measured policy. Payments follow three milestones: the twin agrees across engines (week 2), the policy runs on your robot's computer at its target rate (week 6), and the final evaluation and handover (week 10).
- SubscriptionMonthly
- The pipeline keeps running after the sprint: new demos, retraining, regression tests in the twin and fresh evaluation reports. You pay a monthly fee per robot type, and training compute is billed as usage.
The audit and the sprint are fixed price. We quote the sprint after the audit, once the robot, the task and the data are known. Training compute is included in the sprint up to an agreed cap.
The ten weeks, and the check that closes each step.
| Weeks | Work | Done when |
|---|---|---|
| 1–2 | Robot model cleaned up and loaded in two engines. Actuators fitted from your logs. | The two engines agree within an agreed tolerance. |
| 3–4 | Controller trained in the twin, then run on your robot under a harness. Base-model bake-off on your demos. | The controller tracks commands on the real robot. A base model is chosen. |
| 5–6 | Demos expanded in sim and mixed with real ones. VLA fine-tuned and exported to the robot's computer. | The policy runs on the robot at its target rate. |
| 7–9 | Two rounds of real evaluation against your baseline, with failure analysis between them. | The agreed success threshold is met, or the gap is explained. |
| 10 | Weights, pipeline configs and the evaluation report handed over. | Your team reproduces one training run alone. |
Hardware weeks use conservative motions, a harness or gantry, and your safety staff on site.
What is fixed scope, and what is research.
Some layers are well understood, and we price them as fixed scope. Others are open research problems. We take those on as time and materials, and we say so at the audit.
| Layer | Plug-and-play (fixed price) | Custom R&D (time and materials) |
|---|---|---|
| Sim twin | URDF or MJCF import, actuator fitting, system identification and a two-engine check, for arms and supported legged robots. | Scanned workcells, deformable objects and tactile sensing. |
| Controller | Arm impedance or end-effector control, joystick locomotion, a ported whole-body tracker, a skill API and a safety layer. | New actuator types, models learned from real data, dexterous hands. |
| VLA | Data checks, a base-model bake-off, a fine-tune and a TensorRT export to Jetson Thor. | Learning from corrections, long-horizon tasks, moving from a gripper to a hand. |
| World model | A policy evaluator, once you have 20 to 60 hours of robot video. | Synthetic training trajectories and planning inside the model. |
| Evaluation | Confidence intervals, sequential A/B tests, and rank agreement for any sim or world-model claim. | Safety red-teaming inside a world model. |
Why work with us?
You own the weights.
The policy runs on your robot's computer or your own GPUs. Nothing calls a hosted model at run time, and nothing stops you training it again without us.
The base model is picked by testing.
Open VLAs differ a lot on small datasets. We fine-tune each candidate on your demos and keep the best one. We are not tied to any one model.
Sim, controller and evaluation in one engagement.
Fine-tuning a VLA is the cheap step. The twin, the controller under it and the evaluation that proves it works are the hard parts, and they come together.
We are a model-training and evaluation team. We publish our own tests of model claims, with controls, including the claims that fail. Your evaluation report is held to the same standard.
What we need from you.
Your robot's model
A URDF or MJCF file, motor logs, and the camera layout.
Demos of the task
100 to 200 teleoperated demos by week 3, in any common format.
A safety contact
Someone on your side who signs off each hardware test.