Event

UniWALM Unified World-Action-Language Model Released

1 reports · 1 sources

Rahul Sajnani and other researchers propose UniWALM, a robot manipulation model that simultaneously learns video, action, and language within a single shared representation, with each modality serving both as a conditioning signal and as a supervision target; the team also proposes the Flow Matching with Correction (FMC) training objective, which lets the model recover from perturbed action trajectories without adding extra overhead at inference time. On the real-world Find-the-Item task, generated textual memory raises the success rate from 11.1% to 61.1%, while existing state-of-the-art VLA models achieve success rates no higher than chance level (≤33.3%). This work is a 25-page preprint, published on arXiv cs.RO.