Skip to content

Bring the envs README up to date and register all 22 environments - #197

Merged
yichao-liang merged 2 commits into
masterfrom
envs-readme-solve-gifs
Oct 1, 2026
Merged

yichao-liang merged 2 commits into
masterfrom
envs-readme-solve-gifs

Conversation

@yichao-liang

@yichao-liang yichao-liang commented Oct 1, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Brings predicators/envs/README.md up to date, with every claim re-measured on October 1, 2026.

Previews.
Each preview is now the solve GIF the RoboDisco site shows, linked to that environment's page, so the README and the site show the same demos.

All 22 environments.
The Gymnasium registry gains the seven RoboDisco environments it lacked (Balloons, Bridge, BusyBoard, Crane, IceRink, Launcher, Magnets).
All 22 build, reset, step and render through the wrapper with the Quick Start configuration.
The table lists all 22 with the site's descriptions and marks the five benchmark domains.

Columns, measured.

  • Tasks: whether initial states vary across three train and three test tasks. Ants, Circuit and Float were wrongly marked fixed.
  • Skills: whether get_gt_options builds a skill library.
  • Oracle (was Demos): which approach solves the test task with the README command in the default configuration.
    The old column claimed Boil, Domino, Fan, Float and Grow; today the default configuration solves Ants, Balance, Blocks, Circuit, Cover, Crane, Launcher and Magnets with oracle, and Grow with oracle_process_planning.
    Boil and Bridge solve with oracle_process_planning under the settings of their oracle tests.
    The notes say why the rest fail.
    The pre-refactor commit 72d7258ed gives the same results, so the refactor did not cause any of them.

Fixes to stale or broken text.

  • The note that Circuit, Cover, Laser and Switch fail to build was wrong: every environment builds alone. Circuit, Laser and Switch fail only when another environment was built earlier in the same process, and the README now says so.
  • The Standalone API example passed a raw array to env.step, which raises; it now wraps it in Action.
  • New: task selection and seeding through reset(), the observation layout, the benchmark settings file, and the two skill libraries (CFG.skill_library).

The smoke script and the notebook count 22 environments, and the smoke script no longer blames missing CFG for the same-process failures.
The environment bugs found along the way (the same-process failures, and the default-configuration crashes in Domino, Fan, Grow and Coffee) are left for a separate change.

Test plan

  • Every README Python snippet runs (Quick Start, configuration, make overrides, Standalone API).
  • tests/envs/test_robodisco.py passes with 22 environments.
  • scripts/robodisco_getting_started.py builds all 22 environments in one process; only Circuit, Laser and Switch fail, as documented.
  • Oracle column: the README command with both oracle approaches on every environment, on this branch and at 72d7258ed.
  • GitHub's rendered README serves all 22 previews through its image proxy.

🤖 Generated with Claude Code

The Environments table previewed each environment with a random-action
GIF. Each preview is now the solve GIF the RoboDisco site shows, linked
to that environment's page there, and the legend says what it shows.
Re-measured every claim in predicators/envs/README.md on October 1:
- The Gymnasium registry gains the seven RoboDisco environments it
  lacked (Balloons, Bridge, BusyBoard, Crane, IceRink, Launcher,
  Magnets); all 22 build, reset, step and render through the wrapper.
- The table lists all 22 with the site's descriptions; Tasks and Skills
  follow the measured task variation and skill libraries, and the
  Oracle column (was Demos) names the approach that solves the test
  task in the default configuration, with the reasons for the rest.
- The stale note that Circuit, Cover, Laser and Switch fail to build is
  replaced by the actual failure: Circuit, Laser and Switch fail when
  another env was built earlier in the same process.
- The Standalone API example wraps its actions in Action, so it runs.
- New: task selection and seeding through reset(), the observation
  layout, the five benchmark domains, and the two skill libraries.

The smoke script and the notebook count 22 environments, and the smoke
script no longer blames missing CFG for the same-process failures.
@yichao-liang yichao-liang changed the title Show solve GIFs from the RoboDisco site in the envs README Bring the envs README up to date and register all 22 environments Oct 1, 2026
@yichao-liang
yichao-liang merged commit 864ed48 into master Oct 1, 2026
14 checks passed
@yichao-liang
yichao-liang deleted the envs-readme-solve-gifs branch October 1, 2026 11:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant