Can the fine tuned model on the “internmanip” be deployed on a robotic arm

[THTHDGCS]

If I fine-tune a VLA model with InternManip, can I deploy it on a robotic arm? If so, please provide a deployment suggestion, Thank you.

[JiantongChen]

Based on our experience hosting the IROS offline competition, it is indeed possible to use InternManip to control a real robotic arm.

Here’s an overview of the main steps involved:

1. Implement the real-robot evaluator and env modules
   The core development effort lies in adapting the simulation environment’s code to a real-robot setup.
   In particular, the env should be able to:

   • Return observations (obs) from the real robot
   • Accept actions as input to control the robot arm

2. Develop the agent module
   The agent should handle the conversion between the model’s output and the action space expected by the env at each timestep.

3. Run the evaluation following the standard workflow

   • Launch the model and agent as the server
   • Launch the evaluator and env as the client
   • The evaluator sends the real robot’s observations to the server
   • The model processes them, the agent converts the results into actionable data, and the env executes them on the robot arm

A more detailed tutorial will be added to the documentation in the next release.