Testing ACT Model in Isaac Lab Simulation
I've been working on integrating a trained ACT (Action Chunking with Transformers) model with NVIDIA Isaac Lab to test a this dataset. I would like to initiate a discussion about performance evaluation strategies for behavior cloning models in simulation environments.
I trained an ACT model using LeRobot on the NVIDIA PhysicalAI dataset for the “panda-open-cabinet-right” task.
Currently, I am trying to use the available environment with the Franka robot in ISAAC Lab (https://isaac-sim.github.io/IsaacLab/main/source/overview/environments.html).
I have a few questions:
- Does anyone have experience with the integration of Isaac Lab and LeRobot?
- The environment is functioning correctly, and I can control the robot, but I’m struggling to get the trained model to work. I'm encountering errors related to differences in input/output dimensions and channel mismatches when loading the model.
Any insights or suggestions would be greatly appreciated!
Thanks !
you need to collect the data using imitation learning pipeline of isaaclab and then you can use the hdf5 generated to train a policy you want to like ACT
Thank you @sakehaosdfadfasf for your inputs. Yes, that's a good idea to collect data and then train, but then how can I use the dataset that is provided? Is this of no use then? This dataset is collected in a similar way, and I have already fine-tuned ACT and pi0 on this data; now I want to test the performance and evaluate the policy performance. Can you suggest how can I do that?
Thanks!