Recova

“Recova's core thesis is that task execution and failure recovery should be treated as separate, specialized policies rather than expecting a single VLA model to handle both.”

“adding recovery skills on top of DAgger-finetuned task policies raised mean success from 77.5% to 87.5%”

“Even π0.5 — a state-of-the-art VLA from Physical Intelligence/Google — scored 0.0% on LIBERO-Pro across all six settings”

“Three authors are NVIDIA researchers (Bjorck, Yu, Yin, Kautz, Fan, Liu). The paper uses NVIDIA L40 GPUs for training”

“Intel RealSense cameras (4 per workstation: overhead, front, two wrist) are used for all real-robot perception”

“Claude Opus 5.5 system card is referenced (Reference [2]) in the context of language-model agents with multimodal understanding and tool-use capabilities.”

“GPT-6 Astra system card is referenced (Reference [51]) in the same context.”

“Isabella Liu, UC San Diego, lead author.”

“Yuke Zhu, UT Austin / NVIDIA. Prolific researcher in robot learning, co-author of LIBERO benchmark”