Open Diffusion VLA Jocasta-diff-VLA Released for Robot Control

The RobotLearningLab team released Jocasta-diff-VLA — a LoRA adaptation of Dream-VLA-7B with a Flow Matching action head for controlling a Franka Panda manipulator in robosuite, achieving a 64.8% success rate on 512 test scenes and including a checkpoint, fine-tuning code, and an independent audit of all episodes.

Compiled by Sergey KostenchukPublished 2026-11-04Updated 2026-11-04

2026-11-04 Research HuggingFace

🤖 Open Diffusion VLA for Robot Control Released

RobotLearningLab has released Jocasta-diff-VLA — a LoRA adaptation of Dream-VLA-7B for a simulated Franka Panda manipulator in robosuite. Fine-tuning on 8,140 demonstrations with a Flow Matching action head: 64.8% success rate on 512 test scenes, compared to 0 out of 64 for the base model on PickPlace.

🌍 The entire verification suite is open: checkpoint, fine-tuning code, conversion of predictions to controller commands, and an independent audit of all 512 episodes. Such transparency is rare in robotics, making the verification of others' results inexpensive.

👤 The model is available on Hugging Face, the code is on GitHub, and inference runs on a single CUDA GPU. Limitations: currently simulation only, single seed, weak ToolHang performance (9 out of 64), and transfer to real robots is unconfirmed.

Source 1: https://huggingface.co/RobotLearningLab/Jocasta-diff-VLA Source 2: https://github.com/alanrbtx/Diffusion-VLA-tuning

Sources