Vrushabh27/phoenix-molmospaces-eval

Phoenix — MolmoSpaces Benchmark Evaluation Artifacts

Complete evaluation artifacts for Phoenix, the first fluid model for physical intelligence developed by Geodesic, evaluated on the MolmoSpaces benchmark suite (Franka arm, joint-position action space). Evaluations run June–July 2026. Policy code is not open-sourced; evaluation-harness patches used are published at https://github.com/Vrushabh27/phoenix (patches/).

Results (oracle success condition, official eval_to_csv.py)

Benchmark	Episodes	Success rate
Pick-v1.1 (MS-Pick, bench-v1)	1000	73.7% — leaderboard entry allenai/molmospaces#137
Pick-v1.5 (bench-v2)	1000	72.6%
Pick-v2-classic (hard bench)	1000	58.6%
Pick-v2-filament	1000	56.8%
Pick-v2-rand-cam	1000	58.1%
Pick&Place-v1	1000	43.8%
Pick&Place-v2	1000	44.4%
Pick&Place-color-v2	1000	43.9%
Pick&Place-NextTo-v2	1000	10.8%
Open-v1	1000	37.3%
Close-v1	915	49.95%

Multi-benchmark leaderboard entry: allenai/molmospaces#144.

Repository layout

<benchmark>/                  e.g. classic/, filament/, nextto/, pick_v15/, ...
  videos/ep<NNNNN>/           per-episode mp4s (all recorded cameras, RGB + depth)
  manifest.json               episode index -> exact (h5 file, trajectory) provenance
  Phoenix_<benchmark>_oracle.csv   official per-category score sheet
<benchmark>/pods/<pod>/       raw eval_output + logs pushed directly from eval machines
pick_v15/, pick_v4*/          earlier single-machine runs (Pick-v1.5, Pick-v1.1)

Guarantees: each benchmark's manifest.json maps every benchmark episode (0..N-1) to exactly one clean evaluated instance — full coverage, zero duplicates. Episodes affected by external API outages were quarantined and re-run; scores are computed only over the manifest-selected set using MolmoSpaces' own scorer.

Contact: Vrushabh27 (HF) / Geodesic. See also leaderboard issue allenai/molmospaces#137.