Frequently Asked Questions

What should I run first?

Use PressButton-Am-EE-Delta-PID-Direct-v0 with one environment. It exercises task registration, scene creation, reset, observations, actions, and stepping without Pyroki or acados.

Why are the product and package names different?

AM-Bench is the project and benchmark name. am_isaac and am_isaac_il are the repository's Python package names and therefore appear in imports and filesystem paths.

Which operating systems are supported?

The maintained path is native Linux with an NVIDIA GPU, Isaac Sim 5.1, Isaac Lab, and Python 3.11. WSL, macOS, and CPU-only simulation are outside the maintained runtime boundary. Remote or containerized setups must provide compatible NVIDIA graphics and Vulkan access.

Why use the EE oracle?

The EE-only profile isolates task geometry, scripted behavior, observation/action wiring, and learned-policy behavior from mobile-manipulator IK and base control. It is a diagnostic baseline, not a physical aerial robot.

Use an EE-target interface unless the experiment specifically studies direct base-plus-joint commands. The paper's representative ablation found better task success and lower limit use for the EE-target path, especially with whole-body MPC, but that result is an experimental finding rather than a guarantee for every task.

Are aerodynamic effects always enabled?

No. Environment configs expose saturation, aerodynamic effects, wind, action noise, and observation noise independently. The learned-policy evaluator's --disturbance option enables saturation, aerodynamic effects, and wind together for that run.

Where should datasets and checkpoints live?

Use storage outside version control. Recorder defaults are convenient for local development, but datasets/, checkpoints, videos, and evaluation outputs are generated artifacts and are not part of the public source release.

Can I replay a canonical dataset exactly in simulation?

Not through a maintained public entry point. Exact replay also requires reset and scene state beyond the policy dataset. The supported gate is static dataset validation plus inspection of recorded videos.

Why does OpenPI use a separate server?

The maintained OpenPI stack has its own pinned dependencies and serves policy inference over a network connection. Isaac Sim remains in the AM-Bench environment while the OpenPI process loads and serves the checkpoint.

How should I compare results?

Match the full evaluation configuration: task ID, checkpoint and action semantics, observations, policy rate, seed policy, rollout count, episode length, disturbance flags, and controller profile. Use both task outcomes and tracking/limit diagnostics.