smolvla-metaworld

A vision-language-action model solves robotic manipulation tasks by interpreting natural language descriptions and executing corresponding actions on a robot. This Skill implements such a policy finetuned on the MetaWorld benchmark for autonomous manipulation.

The MetaWorld MT50 benchmark includes 50 diverse robotic manipulation tasks finetuned on a Rethink Sawyer arm. Applications cover robotic assembly, object manipulation, and autonomous manipulation in simulation environments.

Yes, you can run vision-language-action policies on a Rethink Sawyer robot. This Skill requires a robot HAL with a Sawyer embodiment and an RGB camera to execute the manipulation tasks in simulation.

You execute natural language commands for autonomous robot manipulation by running this Skill with the OpenRAL command line interface. It interprets the language descriptions and translates them into robotic actions on the Sawyer arm.

This robotic manipulation Skill requires a robot HAL with a Rethink Sawyer embodiment and an RGB camera. It runs the MT10 or MT50 suites and includes 5 demo scenes for evaluation in simulation.

Limitations of using a vision-language-action policy for robotic assembly include dependency on a specific Rethink Sawyer embodiment and an RGB camera. The policy is finetuned for the MetaWorld MT50 benchmark tasks in simulation.