RoboBRIDGE

“On RoboCasa, average success rose from 3.7% to 7.5% (and 6.2% to 11.4% excluding pick-and-place), across three different VLA backbones (Section IV-B).”

“Table III shows monitoring alone adds +8.1 percentage points with Claude Opus 4.6 as the diagnosis backbone”

“GPT-5 mini showed +6.2% gain with monitoring.”

“Gemini-3 Flash showed only marginal improvement, suggesting insufficient reasoning capacity for failure diagnosis.”

“RoboBRIDGE: A Modular Framework for Bridging Policies to Robust Real-World Robotic Agents”

“The paper directly challenges the dominant industry assumption that scaling VLA models (more data, more parameters) will yield deployable robots.”

“Franka Emika — Franka Emika Research 3 arm used in both RoboCasa simulation and real-world deployment. Hardware platform for validation.”

“Universal Robots — UR7e used as a second real-world robot platform, demonstrating cross-embodiment transfer.”

“Omron — Mobile base mounted with Franka arm in RoboCasa simulation setup.”

“LangChain — Referenced as an example of general-purpose LLM orchestration frameworks that inspired the RoboBRIDGE design analogy.”