RobotAgentBench: hybrid Codex-InternVLA-A1.5 system averages 48.5 on household agent tasks
jiqizhixin · x · 2026-11-14
A preprint from Aether AI, UC San Diego, and UC Davis introduces RobotAgentBench, which evaluates robot agents beyond executing explicit instructions—probing task discovery, contextual reasoning, failure recovery, and robustness to disturbances.
- A real-to-sim pipeline builds 50 home layouts with 90 atomic, 18 short-horizon, and 8 long-horizon tasks plus 4 open-ended scenario types.
- A shared planning–execution–verification structure compares learned policies, code-based control, and hybrid systems; the hybrid can use code to position a gripper before invoking a learned grasping policy.
- The hybrid Codex–InternVLA-A1.5 system reportedly reaches a 48.5 task-level average and a 15.4 open-scenario score.
- Scores are benchmark-specific, not universal household success rates; a matched real-home subset shows lower absolute performance.
More from Embodied
- Ropedia open-sources Xperience 2: 1,589 3D scene scans and 131 hours of egocentric demos for physical AI — liuziwei7 · 2026-11-14
- Huawei Noah wins BEHAVIOR-1K Challenge with 48% task success rate, 4x last year — 机器之心 · 2026-11-14
- ASU IRIS Lab recruiting fully funded PhD students for Fall 2027 in contact-rich dexterity — YuXiang_IRVL · 2026-11-14
- CoRL 2026: a robot shot my Polaroid, a Cybercab drove me around — what evidence would convince you? — bousmalis · 2026-11-14
- One engineer builds complete AI smart glasses from PCB to finished product — IgorCarron · 2026-11-14
- Robotics lab humor: strap the spare motors to a cart and see if it flies — GoddenThomas · 2026-11-14