Robot adaptation work is getting more grounded in contact and instruction control

MoSS makes the day's strongest empirical case for adding physical feedback directly into vision-language-action models. It keeps tactile and torque inputs in separate streams, then lets them interact with the action model through shared attention. On four real robot tasks, the full model lifts GR00T N1.5 from 20.8% average success to 49.0%, and pi_0 from 26.1% to 45.9%. The reported overhead is small at 1.11x for the dual-signal setup. The task mix matters here: cup unstacking, egg pick-and-place, board erase, and plug insertion all depend on contact cues that vision alone can miss.

DeLock focuses on a narrower but important failure mode: low-data post-training can make a robot stop obeying new instructions after it learns a task. The paper keeps the visual encoder close to its pretrained state during fine-tuning, then uses Contrastive Prompt Guidance (CPG) at test time to bias action generation toward a new prompt. In 8 tasks with 20 trials each, it posts 19/20 on two concept-lock tests and 11/20 to 14/20 on several spatial-lock tests. The gains over the low-data RETAIN baseline are large, including 19/20 versus 0/20 on T2 and 13/20 versus 1/20 on T8. Ablations also show that CPG is carrying much of the spatial generalization load.