YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
EFE-Coding: Embodied First Exam workspace (Claude Code)
Scripted privileged-mode solvers for the Embodied First Exam, runnable without docker, plus collected trajectories and a by-setting mirror of the public leaderboard runs.
Deploy
tar xzf efe_workspace.tar.gz && sh efe_workspace/bootstrap.sh # ~15 min, downloads ~21 GB of RoboCasa assets
. ~/efe/env_robocasa.sh
python3.11 ~/efe/solvers_src/robocasa/run.py robocasa-open-drawer-i00-privileged
bootstrap.sh clones the benchmark repo at commit 5e7c183, installs the pinned Python packages into
~/efe/py-robocasa (PYTHONUSERBASE, nothing system-wide), and fetches RoboCasa @756598a plus its assets from the
original hosts with the SHA-256 checks of the benchmark's own Dockerfile. The assets are not re-hosted here.
Contents
efe_workspace.tar.gz:bootstrap.sh,env_robocasa.sh,setup_robocasa.sh,solvers_src/robocasa/(casa_ctl.py: IK, collision-checked grasp planning, base moves;fam_*.py: one solver per task family;run.py: solve, replay twice, write the deliverable) andsolvers/<task>/for every solved task.collected_trajectories/claude_robocasa/<task>/:trajectory.npz,gen.py,NOTES.md,result.json(17 tasks, see status below).collected_trajectories/gemini_antigravity/efe_successful_trajectories_2026-11-19.zip: the Antigravity (Gemini) sessions' successful trials (341 Harbor-verified + 65 replay-only).leaderboard_runs/by_setting/<harness>-<model>-<effort>.zip(+_manifest.csv,summary.json): a full mirror ofMLL-Lab/embodied-exam-runs, one zip per leaderboard setting, 17 settings, 16,553 runs, 0 missing files, ~162 GB. Each zip holds<setting>/<suite>/<run>/{run.json, trajectory.jsonl.gz, steps.bin, replay.mp4, poster.jpg}(files that the source does not have for a run are absent). Every zip's sha256 was checked against the LFS pointer after upload.leaderboard_runs/<suite>/: older per-suite zips of the same data (superseded byby_setting/).
Status (2026-11-19): RoboCasa privileged, 17 of 29 solved
drawers (open, close), doors (open/close single and double), microwave start/stop, coffee button, sink faucet on/off, sink spout, stove knob on/off; pick-and-place: counter-to-stove, sink-to-counter, microwave-to-counter. Not yet: stove-to-counter, cab-to-counter, counter-to-cab, counter-to-sink and the composite tasks (coffee serve/setup mug, microwave thawing, prepare coffee, pre-soak pan, restock pantry, arrange vegetables).
Verification is casa_tool.Session.replay run twice with identical state hashes (the call the task text says the
verifier makes). None of these was run through Harbor. Solvers read simulator state (allowed in privileged mode) and
never read solution/ or tests/.