multitask_7tasks_GR00T17_40k

GR00T N1.7 policy trained with LeRobot on 7 manipulation tasks. Final checkpoint, 40,000 steps, batch size 64.

Fine-tuned from nvidia/GR00T-N1.7-3B on the private dataset TAPTIC/multitask_7tasks_reduced: 550 episodes, 186,419 frames at 10 fps, with 10% of episodes held out for evaluation. The state is 40-dimensional: raw tactile taxel readings were removed.

Task	Episodes
Take the elliptical cylinder and put it in the corresponding hole.	100
Take the credit card and insert it in the wallet.	100
Take the fruits and put them in the basket, one by one, without dropping them.	100
Pick up one cup from the stack of cups.	100
Open the bottle, unscrew the cap.	50
Gently grasp the plastic glass and move it to the target spot.	50
Clear the table by putting the objects into the box.	50

License

Derived from nvidia/GR00T-N1.7-3B. The base model repository includes the NVIDIA License, copied here as LICENSE, which limits use to non-commercial research or evaluation. Its README and the Isaac-GR00T repository name the NVIDIA Open Model License for the weights. This model is distributed for non-commercial research and evaluation only.