Hi guys,
First off nice work, the V+L→A framing (dropping the LLM backbone and using DINOv3 + BERT directly instead) is a super nice change from the usual "stack a bigger VLM" trend, especially with the RTX 4090 / <1GB VRAM numbers.
I contribute to LeRobot and I'm planning to port TurboVLA in as a lerobot_policy_turbovla plugin (LeRobot's out-of-tree policy mechanism — no core PR needed to use it, just a pip-installable package that auto-registers).
Wanted to open this to:
- give a heads-up in case it's useful context on your end (adoption signal, or if anyone else is already doing something similar),
- ask if there's a preferred way to reference/cite the released checkpoints (
H-EmbodVis/TurboVLA on the Hub) once I re-publish a converted version in LeRobot's checkpoint format,
- flag that the plugin will initially carry over the sim-only training scope (LIBERO / RoboTwin) as-is, and note that clearly rather than imply real-robot validation that isn't there.
Tracking issue on the LeRobot side, for anyone who wants the details: huggingface/lerobot#4400
No action needed from you, just didn't want to do this silently.
Happy to link back here once there's something runnable.
Hi guys,
First off nice work, the V+L→A framing (dropping the LLM backbone and using DINOv3 + BERT directly instead) is a super nice change from the usual "stack a bigger VLM" trend, especially with the RTX 4090 / <1GB VRAM numbers.
I contribute to LeRobot and I'm planning to port TurboVLA in as a
lerobot_policy_turbovlaplugin (LeRobot's out-of-tree policy mechanism — no core PR needed to use it, just a pip-installable package that auto-registers).Wanted to open this to:
H-EmbodVis/TurboVLAon the Hub) once I re-publish a converted version in LeRobot's checkpoint format,Tracking issue on the LeRobot side, for anyone who wants the details: huggingface/lerobot#4400
No action needed from you, just didn't want to do this silently.
Happy to link back here once there's something runnable.