Robots controlled by generalist AI outperform specialists without task-specific training

The authors report stronger benchmark and real-robot performance when generalist models directly revise robot actions, rather than relying on policies trained on demonstrations.

Big Tech

Chen Chen · Omid Bagherieh · Naveen Nikamanth Appiah Balaji · Jesse T. Gonzalez · Negin Heravi · Fahad Islam · +7 more

Apple

Research Digest·9 November 2026·2 min read

Illustration: Robots controlled by generalist AI outperform specialists without task-specific training

In one line

Generalist AI models control manipulators more reliably than specialist policies, with 85.6% zero-shot success on RoboLab-120.

Before you read this

Two ideas this paper builds on

agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.173 papers →agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.223 papers →

Read it as

Robot manipulation usually requires many demonstrations. Specialist policies that map camera images to actions are brittle and fail often.

This work introduces Agentic Manipulation Policies (AMPs), where a generalist AI model directly controls the robot and adapts as the task unfolds. The AMP1 harness works on various robots without task-specific training.

AMP1 achieved 85.6% success on benchmarks without demonstrations, far exceeding specialist policies. With persistent memory, it reduced failure rates 4.6 times on two-arm tasks. This shows generalist AI can handle manipulation robustly, offering an alternative to specialist training.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agentic manipulation policy agent harness zero-shot learning agent memory emergent behavior few-shot learning