TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

benchmarks evals

Aerial Manipulation Got a Systems Benchmark

AM-Bench varies aircraft embodiment, controller, disturbance and policy instead of scoring only the final learned behavior.

Published Updated Story ID: mp-2026-09-02-018
Read the complete editionStory JSON

Summary

AM-Bench varies aircraft embodiment, controller, disturbance and policy instead of scoring only the final learned behavior.

The simulation suite includes underactuated, fully actuated and overactuated multirotors across 12 contact, transport and constrained-interaction tasks. Configurable wind effects, actuator saturation and standard controllers make it possible to isolate system interactions. Three simulation studies, modeled-effect checks and a hardware pipeline test establish diagnostic coverage rather than a leaderboard winner.

Why it matters

AM-Bench varies aircraft embodiment, controller, disturbance and policy instead of scoring only the final learned behavior.

Limits and context

No additional limitation was separately recorded.

Key claims

  1. AM-Bench varies aircraft embodiment, controller, disturbance and policy instead of scoring only the final learned behavior.

    Evidence: source-2026-09-02-020

Sources

  1. arXiv preprint 2609.00641arXiv · primary research

Corrections

No corrections have been recorded for this story.