| tags: | |
| - robotics | |
| - vision-language-action | |
| - vla-arena | |
| - paimon | |
| pipeline_tag: robotics | |
| # Paimon | |
| This repository contains the evaluation results for **Paimon**, a | |
| vision-language-action policy evaluated on the VLA-Arena benchmark. This is a | |
| results-only submission for the checkpoint at training step **40,000**。 | |
| The evaluated system uses a two-stage pipeline: a VLM-Agent first generates or | |
| decomposes the subtask, and the VLA then executes the resulting subtask in the | |
| simulator. | |
| ## VLA-Arena evaluation | |
| Paimon was evaluated on all **33 VLA-Arena task/level cells** using the | |
| official-style evaluation setup. | |
| - Episodes: **5,100** | |
| - Successes: **3,767** | |
| - Episode-weighted success rate: **73.86%** | |
| - Cell-mean success rate: **73.08%** | |
| - L0 / L1 / L2 cell-mean success rate: **90.58% / 71.39% / 57.27%** | |
| The cell-mean rate is the equal-weight average of the 33 L0/L1/L2 success | |
| rates in `results/paimon-arena-agent-ft-step40000.json`. | |