AM-Bench¶
A modular simulation suite and benchmark for aerial manipulation policy learning.
AM-Bench studies aerial manipulation as a system-level problem. A policy does not act in isolation: its behavior depends on the robot embodiment, the action interface, the low-level controller, and the physical effects acting on the platform. The suite makes those choices configurable under shared tasks and evaluation outputs.
12 tasks
Instantaneous interaction, object transport, articulated objects, and constrained contact.
Five embodiments
Four multirotor-based platforms spanning underactuated through overactuated designs, plus an EE-only oracle.
Common policy boundary
Scripted policies, ACT, Diffusion Policy, PI0, and PI0.5 share canonical data and evaluation conventions.
Dynamics-critical evaluation
Control allocation, rotor saturation, actuator dynamics, drag, wind, ground effect, and near-wall effect.
Start with one working environment¶
Install AM-Bench beside an Isaac Lab checkout, activate the Isaac Lab Python environment, and list the registered environments:
source ../IsaacLab/env_isaaclab/bin/activate
uv pip install -e source/am_isaac
uv pip install -e source/am_isaac_il
python scripts/environments/list_envs.py
Continue with Installation for prerequisites and optional controller dependencies, then run the bounded check in First Simulation.
What the benchmark is for¶
AM-Bench is designed for controlled comparisons across five axes:
- task environment;
- robot platform;
- high-level policy and action interface;
- low-level control;
- disturbances, actuator dynamics, and saturation.
It supports questions such as whether an end-effector target is a better policy interface than direct base-plus-joint targets, how controller choice changes policy outcomes, and where an embodiment reaches its physical limits. See Benchmark Design for the system decomposition and Experimental Findings for the paper snapshot.
Choose your path¶
- Run a task: First Simulation
- Collect data: Collect Demonstrations
- Train a baseline: Policy Guides
- Interpret an environment ID: Environment IDs
- Add a benchmark component: Extend AM-Bench
- Cite the project: Paper and Citation