Roboticz Lab ES
‹ News

Mobile robots, demonstrations and action models | AI robotics

Three November 2 preprints test a complete mobile robot, learning from human demonstrations and a filter for model-generated actions.

  • mobile manipulation
  • robot learning
  • world action models
  • human demonstrations
  • AI
  • news
Generated image: macro detail of lab instrumentation, accompanying the entry Mobile robots, demonstrations and action models

What has actually changed in the last 24 hours? Three preprints make the same useful point: a robot does not improve just because its model is bigger. It needs a body, data, and a way to reject bad actions before motors move. These are author-reported results, not products or independent validation.

Pixel treats the robot, teleoperation and control as one system

Pixel, published on November 2, introduces a wheeled humanoid with compliant grippers, teleoperation interfaces and whole-body control. Its authors evaluate the stack on tasks combining navigation, active perception and manipulation.

The useful part is not a low-cost claim, because the preprint does not provide a price that lets us compare it. It is the decision to build hardware, demonstration collection and the policy as one chain. For a home project, the lesson is simple: before swapping models, make sure the camera, controller, actuators and data logger share the same timing reference.

Pixel-UMI addresses the gap between a human demonstration and a robot

The same team also released Pixel-UMI. It turns human head and gripper trajectories plus reconstructed scene geometry into targets for the robot base and joints. The pipeline also uses human corrections during data collection.

The authors report 87.0% average task progress across five tasks, including bimanual manipulation and reaching in constrained spaces. That is their own evaluation. Still, it makes a practical point: copying human positions without accounting for furniture, cameras and robot delay often leaves you with a convincing recording and clumsy execution.

If you record demonstrations, begin with one repeatable task and save the scene, the command sent and the time when the robot actually moved. You do not need a humanoid to apply that discipline.

WISE checks whether an imagined future agrees with an action

The third preprint, WISE, does not train another action model. It samples several proposals, checks whether each action agrees with the future motion imagined by that same model, then executes the highest-scoring option. With four candidates per decision, the authors report RoboLab success rising from 6.0% to 24.0% with Cosmos 3 Edge. On LIBERO, they report an increase from 75.1% to 78.3% with DreamZero.

That is not a guarantee for a real arm. These are author-reported benchmark results. But the rule transfers well to a smaller build: if your policy can suggest several trajectories, add a cheap check before sending one to a motor. Joint limits, obstacle distance and an estimate of whether the gripper can still see the object are already a start. A small filter like that often teaches you more than asking the model to be smarter.

Sources
  1. Pixel, arXiv preprint
  2. Pixel-UMI, arXiv preprint
  3. WISE for world action models, arXiv preprint

Every claim above comes from here. If something has no source, it's not in.

Now go build it

Guides with the bill of materials and the logic explained.

Subscribe by RSS