Relationship to FineVLA: fine-grained annotations and language-guided policy learning
Hi FineART team,
Thanks for releasing FineART, the policy, and the training code! Providing dense, human-written subtask annotations for long-horizon bimanual manipulation is a valuable contribution to the community.
I’m the first author of FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies, first released in May 2026:
I noticed a connection between our work and FineART around using fine-grained, action-aligned language supervision to improve robot instruction following.
FineVLA builds process-level annotations for existing robot datasets through a VLM-assisted, human-verified pipeline. Its annotation schema covers ten execution-related dimensions, including the active arm, contact and approach, trajectory and orientation, and failure and recovery. We also study how mixing these fine-grained instructions with coarse task-level labels affects VLA performance and steerability.
From my reading, FineART places particular emphasis on human-written, temporally segmented subtasks throughout long-horizon demonstrations, together with a policy that predicts the next subtask and conditions action execution on it. It would be interesting to discuss how this formulation relates to FineVLA’s execution-level supervision.
I couldn’t find FineVLA mentioned in the current paper. Would you be open to citing it as related work and adding a brief qualitative comparison? In particular, clarifying the differences in annotation granularity, temporal alignment, and how language supervision is used during policy training and inference would help readers understand the respective contributions.
Happy to share further details or discuss the annotation schemas. Thanks again for making this work available!
Best,
Xintong Hu