Apple paper: generalist AI model drives robot arm zero-shot at 85.6%, crushing specialist VLA policies
rohanpaul_ai · x · 2026-11-10
Apple's new paper On Agentic Manipulation Policies shows a generalist frontier model directly commanding a robot arm beats specialist policies trained on demonstrations — with no demonstrations or task-specific training.
Key points:
- The reference harness AMP1 gives models like Claude Opus 5.5 or Gemini camera views, code execution, and direct motion commands in a loop, with no built-in grasping or placing skills — the agent invents its own strategies (searching for hidden objects, using objects as tools, calibrating its cameras).
- On RoboLab-120, AMP1 hits 85.6% success zero-shot vs 49.2% for the best VLA and 42.9% for the best world-action model.
- On a real Franka arm it succeeds in 104 of 108 trials across twelve tasks, cutting failure rates 20x and 15x vs strong baselines.
- Adding a persistent memory store lowers failure rates 4.6x on RoboDojo bimanual tasks.
Main drawback is speed: the robot spends much of each episode waiting on model inference. The authors argue AMPs are substantially more general and robust, offering a promising foundation for manipulation.
More from Embodied
- Developer uses a 1.9B decision model on a laptop to flag dirty restaurant tables in real time — JosephJacks_ · 2026-11-10
- Fable 5.2 beats specialized robotics AI on dual-arm RoboDojo benchmark — FinanceYF5 · 2026-11-10
- LingBot-Touch 开源触觉基础模型：近 5000 小时真实数据预训练 — dair_ai · 2026-11-10
- Norway runs 2,000+ underwater AI laser robots delousing 50 million salmon around the clock — cephaloform · 2026-11-10
- Anthropic robotics model drop: bitter lesson may hold in robotics — beffjezos · 2026-11-10
- Rider shares first-person Cybercab experience: skipping downtown driving in Austin — yunta_tsai · 2026-11-10