Skip to content

Trained models

Models trained end to end with urdf2vla, from the robot's URDF to a GR00T N1.7 checkpoint, and evaluated in Isaac Sim. Each one is on Hugging Face with the files urdf2vla needs to run it.

Robot Model Training data Success in simulation Trained from
FANUC CRX-3iA + Robotiq 2F-85 umegan/urdf2vla-crx3ia-gr00t-n1.7 umegan/urdf2vla-crx3ia-pickplace 20/20 nvidia/GR00T-N1.7-DROID, 50 scripted demos
Franka Panda (its own hand) umegan/urdf2vla-franka-gr00t-n1.7 umegan/urdf2vla-franka-pickplace 19/20 nvidia/GR00T-N1.7-DROID, 50 scripted demos

All were trained and evaluated on one NVIDIA L40S; see Run on NVIDIA Brev for every measured step. They are simulation results only: one scene, one object at a time, fixed lighting.

FANUC CRX-3iA

The trained model picking an apple, a color cube, a lime and a pear (four evaluation episodes, one per object). Each frame shows what the model sees: the wrist camera, then the left and right cameras.

Four moments of one episode: start, grasped, over the bin, placed

Hugging Face umegan/urdf2vla-crx3ia-gr00t-n1.7 (NVIDIA Open Model License)
Robot FANUC CRX-3iA (6 joints) with a Robotiq 2F-85 from NVIDIA's SimReady assets (example)
Cameras wrist + left + right, 256×256 each
State / action 6 joint positions + gripper; 16 actions per call at 20 Hz
Training data 50 scripted demos (50 of 59 attempts succeeded, 33 min): umegan/urdf2vla-crx3ia-pickplace
Training 10,000 steps, batch 32, about 2 h 40 min on an L40S (41.9 GB)
Evaluation 20/20 (apple 5/5, cube 7/7, lime 3/3, pear 5/5); median 174 steps (8.7 s), placed 15 mm from the target
Release check 10/10

The same demos in GR00T's DROID layout (wrist + one exterior camera) scored 19/20: one pear was never released.

Franka Panda

The trained model picking an apple, a color cube, a lime and a pear (four evaluation episodes, one per object): wrist camera, left camera, right camera.

Four moments of one episode: start, picking up, over the bin, placed

Hugging Face umegan/urdf2vla-franka-gr00t-n1.7 (NVIDIA Open Model License)
Robot Franka Emika Panda (7 joints) with the Franka Hand from its URDF (tutorial)
Cameras wrist + left + right, 256×256 each
State / action 7 joint positions + gripper; 16 actions per call at 20 Hz
Training data 50 scripted demos (50 of 63 attempts succeeded, 30 min): umegan/urdf2vla-franka-pickplace
Training 10,000 steps, batch 32, 2 h 42 min on an L40S (41.6 GB)
Evaluation 19/20 (apple 5/5, cube 7/7, lime 3/3, pear 4/5); median 194 steps (9.7 s), placed 18 mm from the target
Release check 10/10

The one failure closed on a pear without lifting it (missed_grasp).

Use a model

Set up the project for the same robot (the example in examples/), download the model next to it, and evaluate:

hf download umegan/urdf2vla-crx3ia-gr00t-n1.7 --local-dir projects/crx3ia/training/hf_crx3ia
mkdir -p projects/crx3ia/datasets/demos_gr00t
cp projects/crx3ia/training/hf_crx3ia/urdf2vla_modality_crx3ia.py projects/crx3ia/datasets/demos_gr00t/
urdf2vla evaluate_vla --project projects/crx3ia --model training/hf_crx3ia --episodes 20 --video

evaluate_vla registers the model's embodiment from the modality config in datasets/demos_gr00t/, which is why the file is copied there. The model card on Hugging Face has the inputs, outputs and training details.

To train it again, or to train your own variant, download the dataset into the project and run train_vla:

hf download umegan/urdf2vla-franka-pickplace --repo-type dataset --exclude "raw/*" --local-dir projects/franka/datasets/demos_gr00t
urdf2vla train_vla --project projects/franka --dataset demos

Each dataset also has the raw recording (raw/demos.hdf5) for converting it again, for example with --layout droid.