ManiSkill Panda StackCube Python recipe: robotics skill example

Grasp one cube and stack it on another with the official Panda motion planner.

Markdown | JSON example and evidence | Canonical contract | All examples

Task and environment

maniskill | 3.0.0b22 | panda | pd_joint_pos | linux | 3.10

Example input

{
  "task": "StackCube-v1",
  "seed": 0,
  "robot": "panda",
  "controller": "pd_joint_pos",
  "obs_mode": "state",
  "sim_backend": "cpu",
  "render_backend": "none",
  "max_episode_steps": 1000
}

Retrieve without installation

curl -fsSL 'https://api.superfastrobotics.com/v1/skills/maniskill-panda-stack-cube?hub_commit=ce1d44c01f1a9f707f055da0888068fe32390e62&record_version=1.1.0&include=experience'

Inspect source, license, environment and execution evidence before preparing. Preparation and execution are separate actions.

Success condition

StackCube-v1 env info.success is true after the upstream solver finishes, truncated is false, and 0 < elapsed_steps < max_episode_steps=1000; only then status is task_complete and exit is 0. The 1000-step motion-planning horizon is a Hub recipe condition, not the upstream default benchmark horizon.

Recorded evidence

task_complete: StackCube-v1 seed 0 completed in the explicit Hub 1000-step motion-planning episode: success=true, terminated=true, truncated=false, elapsed_steps=107, exit=0 and no process timeout. One invocation was made for this task in the LLM-selected revalidation run. This verifies only the recorded CPU/state configuration and one seed, not the upstream default 50-step benchmark, hardware or camera modes.

Source commit: 33967b9e3ead1f841eec57cc9f31d0d8b8cf0907. Record version: 1.1.0. This page performs no execution.