Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision
The paper proposes training a compact “workspace token” so robots can use memory at deployment without live VLM reasoning.
The method uses a VLM during training to identify task-relevant current and past information, then distills that into a lightweight latent memory. The authors say this avoids conditioning policies on full histories, which can introduce spurious correlations and hurt performance. In simulation and hardware tests, the workspace token is presented as a drop-in replacement for observations on memory-heavy manipulation tasks. The paper reports that the token is not just cheaper at deployment, but also improves policy performance. ArXiv · AI/CL/LG's note
The method uses a VLM during training to identify task-relevant current and past information, then distills that into a lightweight latent memory. The authors say this avoids conditioning policies on full histories, which can introduce spurious correlations and hurt performance. In simulation and hardware tests, the workspace token is presented as a drop-in replacement for observations on memory-heavy manipulation tasks. The paper reports that the token is not just cheaper at deployment, but also improves policy performance. ArXiv · AI/CL/LG's note
score 5