RoboARCade launches: agents directly control robots, best GPT-6 setup tops out at 49.8% success
leto__jean · x · 2026-10-13
Researchers released RoboARCade, a unified, modular and extensible benchmark for agentic robot control, built on seven existing benchmarks and made harder via task combinations and greater object pose variations. Paper, code, assets and all agent trajectories (tens of thousands of dollars' worth) are open today.
Unlike hierarchical VLA models or Code-as-Policies, agents here send low-level actions (e.g. end-effector movements) directly to the robot controller, so outcomes reflect the model's own embodied reasoning. Leaderboard highlights:
- GPT-6 Astra (xhigh) + Codex: 49.8% success with the VIA visual interface, 48.3% with Multi-view RGBD
- Claude Fable 5.1: 43.6% (VIA) / 36.8% (RGBD), at $55.50/$33.16 API cost per trial
- Claude Opus 5: 30.3% (VIA); GPT-5.6 Sol: 14.0%
Saturation stays low (mostly 10–36%), showing long-horizon direct motor control remains a major open challenge for frontier models.
More from Embodied
- Dev ships 'EveryDay Carry' firmware for AI Passport, hailed as best network radio — ezshine · 2026-10-13
- Free-Init: scan-free, motion-free initialization for Doppler LiDAR-inertial systems — rsasaki0109 · 2026-10-13
- On-device robot practices soccer throw-ins with no external tracking — chris_j_paxton · 2026-10-13
- GPT-6 Astra and Claude Opus 5 nearly tied on embodied task: 34-6 over 40 votes, 16-point Elo gap — qinzytech · 2026-10-13
- Custom ComfyUI node repurposes Mixamo's 3D rigs to pose 2D characters — Willonfire8 · 2026-10-13
- First Origami Challenge at IROS 2026: 14 teams push dexterous manipulation to the limit — chris_j_paxton · 2026-10-13