Teach the robot, then prove it learned

Imitation learning, vision-language-action models and reinforcement learning, chosen by what fits your hardware and licence, and measured on your arm rather than on someone else’s benchmark.

Start a project How we choose models

A UR robot arm picking a can in simulation — Simulation

The same UR robot arm picking a can on a real table — Real arm

Success in the ACT paper’s own study when the policy predicts chunks of 100 actions instead of one at a time.: 1% → 44%Success in the ACT paper’s own study when the policy predicts chunks of 100 actions instead of one at a time.
What 18 successes in 20 trials really tells you (95% interval). We run enough trials for a number to mean something.: 68–99%What 18 successes in 20 trials really tells you (95% interval). We run enough trials for a number to mean something.
Current VLA models that control force. They output positions, so we add the force limits around them.: 0 of themCurrent VLA models that control force. They output positions, so we add the force limits around them.

What we do

Learned models for arms
Policies trained from demonstrations, and foundation models fine-tuned for your task.
- Teleoperation and demonstration capture
- Imitation learning with ACT and Diffusion Policy
- Fine-tuning VLA models such as π0, GR00T, SmolVLA and OpenVLA
- LeRobot datasets and training pipelines
Reinforcement learning and policy training
Policies that improve through practice in simulation, then transfer to the real cell.
- Reward and curriculum design
- Parallel training in Isaac Lab and MuJoCo
- Domain randomization for sim-to-real transfer
- Human-in-the-loop RL fine-tuning on hardware

Model or no model

What does the robot do when contact starts?

Programmed approach

State machines and behavior trees
Force and torque thresholds
Impedance and compliance control

Learned approach

Imitation learning from demonstrations
Vision-language-action models
Reinforcement learning fine-tuning

How we choose

A learned policy always runs inside a programmed safety envelope: workspace limits, force limits and a fallback behavior. We compare its success rate and cycle time against a scripted baseline, on your arm, before recommending it.

Choosing a learned model

Size, hardware and licence narrow the choice before any training starts.

Model	Size and hardware	Licence and notes
SmolVLA	450M parameters; runs on a laptop	Apache-2.0 code and weights
ACT	About 80M parameters; fast to train	Strong for bimanual tasks from demonstrations
Diffusion Policy	100 network passes per action chunk	Smooth motions, slower to run
GR00T N1	About 3B; 16 GB to run, 40 GB+ to fine-tune	NVIDIA GPU required
OpenVLA	7B; a full fine-tune needs 8 A100s	MIT; LoRA fine-tuning is the practical route
π0 / π0.5	Needs a large NVIDIA GPU	Some weights carry restrictive licence terms