libero-plus

[missTL]

Hello, congratulations to DeepThinkVLA for achieving excellent results on libero-plus. I have some questions to ask about libero-plus.

1. Do you train a single model for all libero tasks, or do you train a separate model for each task?
2. Does the training dataset and the inference/test dataset include libero_90?
3. Which model achieved a score of 79, the SFT model or the RL model?
4. What are your insights on this excellent result? Where does it come from?
   Looking forward to your communication, thank you.

[wadeKeith]

Thank you for your support of deepthinkvla. Our performance on libero plus is zero shot.

1. Our training on libero utilized the official dataset they provided, which means all suites were mixed together, and we only trained one model in total.
2. During the training and testing phases, we aligned with other VLA work, meaning both training and testing were conducted only on object, spatial, goal, and 10.
3. On libero plus, we used the sft model, as it was zero shot, without introducing any further training, including rl.
4. In terms of insights, I believe that when there are many disturbances, the gain brought by CoT is quite good. Additionally, I believe this is not yet the model boundary of deepthinkvla based on a hybrid architecture with reasoning capabilities. We will continue to conduct real machine tests on robochallenge.