FineART-VLA
VLA that predicts its next subtask in language before each action chunk, mid-trained on the FineART dataset (1,718 hours, 151 tasks on a stationary ALOHA) and implemented in LeRobot. Results are self-reported real-robot tasks: spatial instruction following rises from 32.0% to 100.0%.
Scale AI & Hugging Face · Sep 29, 2026 · VLA · built on π0.5Report an error
0Benchmarks reported
Sep 29, 2026Released
OpenWeights
VLAType
Benchmark results0
No results on the tracked benchmarks yet.
Not reported on 15 benchmarks
Wrong or missing number? Report an error