What Is a Visuomotor Policy?

A visuomotor policy is the part of a robot that connects seeing to moving. The Diffusion Policy paper introduces its method as “a new way of generating robot behavior by representing a robot’s visuomotor policy as a conditional denoising diffusion process.”

Firsthand does not train, ship or benchmark visuomotor policies or models. The site states that it captures egocentric video from a head camera and two wrist cameras, depth recovered from the calibrated stereo rig, IMU, 3D hand pose, wrist trajectories, action segments, contact events and a task instruction per episode. Delivery is in RLDS, LeRobot, WebDataset or HDF5. Robot joint or gripper commands, a policy-ready export and policy results are Not stated.

In the Diffusion Policy paper, the policy takes the latest few steps of observation data as input and predicts a run of future actions, some of which are executed before it replans. A visual encoder maps the raw image sequence into a latent embedding.

The paper’s bimanual setup uses scene and wrist cameras plus proprioception, with desired end-effector poses and gripper widths as actions. GR00T N1 (arXiv 2503.14734) describes its model input as robot state, visual observations and a language instruction, and its output as the corresponding motor action.

A vision-language-action (VLA) model also maps images to actions, and adds a language instruction as input. GR00T N1 describes itself as a VLA model that “generates actions from image and language instruction input.” Whether a VLA model counts as a visuomotor policy is a naming choice the cited papers do not settle.

Behavioral cloning describes how a policy is trained, by copying demonstrations, not what the policy is. Diffusion Policy evaluates its policies “under the behavior cloning formulation,” so a visuomotor policy is often trained by behavioral cloning. The second table lays out the three terms.

The cited work says yes with caveats. EgoScale (arXiv 2602.16710) trains a VLA model on over 20,854 hours of action-labeled egocentric human video. Camera motion and hand pose are estimated from head-mounted egocentric RGB by off-the-shelf pipelines. The authors conclude that large-scale human motion provides “a reusable, embodiment-agnostic motor prior.”

GR00T N1 notes that its human video datasets carry no explicit action labels, so it learns latent actions to stand in for them.

Observations paired with actions, in order, on one timeline. Diffusion Policy notes that it inherits a limit of behavior cloning: suboptimal performance with inadequate demonstration data.

Human video has no robot actions in it, so a team derives a target from hand pose or wrist motion, as EgoScale does, or learns latent actions, as GR00T N1 does. A buyer builds the training targets and trains the policy in their own pipeline.

Request a custom dataset and write the tasks, environments, streams, labels and delivery format into the spec. The site states a median of 48 hours from a signed spec to a first validated batch where coverage is already deep, and three to four weeks in thin environments, and billing is per validated hour. If your policy consumes depth, a language instruction or a particular action target, say so in the spec.

What is a visuomotor policy?

A robot policy that maps camera images, usually with the robot’s own state, to actions, learned from demonstrations that pair what the robot saw with what was done.

Is a visuomotor policy the same as a diffusion policy?

No. Diffusion Policy represents a visuomotor policy as a diffusion process, so diffusion is one way to build one.

How is a visuomotor policy different from a VLA model?

A VLA model adds a language instruction to the image input.

Can a visuomotor policy learn from human video?

In the cited work, EgoScale trains on over 20,854 hours of action-labeled egocentric human video, and GR00T N1 learns latent actions for human video that has no action labels.

Does Firsthand train or ship visuomotor policies?

No. The site states no trained, shipped or benchmarked policy or model. Firsthand captures the data, and the buyer trains the policy.

How do I get custom visuomotor training data fast?

Request a custom dataset with the streams, labels and format written into the spec. The site states a median of 48 hours to a first validated batch where coverage is deep, and three to four weeks in thin environments.