Apple's AMP1 paper: general LLMs write code to control robot arms at 96.3% success
APPSO · wechat · 2026-11-11
Apple's November 6 paper AMP1 (Agentic Manipulation Policies) takes a path opposite to mainstream VLA/WAM approaches: a general-purpose LLM watches camera feeds, writes Python to compute trajectories, executes and verifies—no task-specific training data.
Key results: on RoboLab-120 simulation, Gemini 4 Argon + AMP1 hit 85.6% and Claude Opus 5.5 82.8%, vs. 49.2% for the best VLA and 42.9% for the best WAM. On a real Franka arm, error recovery pushed success to 96.3% (vs. 87.0% without), while baselines FLUX3 Action and π0.5 managed 45.4% and 27.8%. With persistent memory, RoboDojo zero-shot success jumped from 26.4% to 84.1%. Model quality matters hugely: removing depth info barely hurt Opus 5.5 but crashed Haiku 5.5 from 51.6% to 10.7%.
Notable emergent behaviors include using a spoon as a tool to reposition an unreachable bottle and self-diagnosing camera calibration errors. The main weakness is speed: 280 seconds per task, 211s spent on inference, limiting use to quasi-static scenes. The article also links this to Apple's HALO humanoid paper on a Unitree G1 driven by Qwen models on a Mac mini, arguing Apple is shifting robotics competition from hardware to software capability.
More from Embodied
- AI lets robotics research thrive in labs with no robotics experts — 4310sy · 2026-11-11
- Tektronix launches ~$500 12-bit DPO scope with AI MCP support in closed beta — IgorCarron · 2026-11-11
- XMAQUINA DAO deploys $1M+ into Sunday Robotics' Memo household robot — xmercury_one · 2026-11-11
- Tesla overtakes Waymo in Texas AV registrations: 1,302 vs 1,297 vehicles — XFreeze · 2026-11-11
- Compass 1 navigation foundation model hits 79.4% success in 320 real-world trials, zero-shot transfers to 4 robot types — abursuc · 2026-11-11
- Towel-folding robot hits 99.5% success, 68% of human throughput on real laundry floor — lukas_m_ziegler · 2026-11-11