DynamicVLA-LIBERO
A VLA model for dynamic object manipulation
2 min read
DynamicVLA is a vision-language-action model for robotic control in tasks involving dynamic objects. This checkpoint is an implementation for the LeRobot library, fine-tuned on all four LIBERO datasets based on a DOM-pretrained model. The model is designed for robotics researchers and developers working on grasping and manipulation tasks. It demonstrates high task execution accuracy across various LIBERO benchmarks by integrating visual and textual data with robot action control.
Key features
- Built on the vision-language-action (VLA) concept for dynamic object manipulation
- Tested on all four LIBERO datasets with an average success rate of 95.5%
- Implemented in the LeRobot ecosystem for easy testing and deployment
- Trained for 40k steps using image augmentation
What you can do
- Evaluate the model policy in the LIBERO simulation environment using the `lerobot-eval` utility
- Test model performance across various task suites, including LIBERO-Spatial, LIBERO-Object, and LIBERO-10
- Integrate the checkpoint into your own robotics pipelines supporting LeRobot and manipulation tasks
Example
Run model evaluation in the simulator using the command:
lerobot-eval \
--policy.path=hzxie/dynamic-vla-LIBERO \
--env.type=libero \
--env.task=libero_spatial \
--eval.batch_size=10 \
--eval.n_episodes=50Read more
DetailsBenchmark success rate: LIBERO-Spatial — 96.6, LIBERO-Object — 98.4, LIBERO-Goal — 93.4, LIBERO-10 — 93.6 (average 95.5)
Training: 40k steps, batch size 72 (3 GPUs), AdamW optimizer, 1k warmup steps
Data: HuggingFaceVLA/libero (1,693 episodes) with image augmentation










Comments