---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- so-101
- evaluation
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# SO-101 push-green-button evaluation rollouts

Merged LeRobot v3 dataset containing 1955 single-arm SO-101 evaluation episodes
(491,385 frames at 20 FPS) from 38 source datasets recorded August 25–September 16, 2026.
All sources share the same state/action schema and `front`, `top`, and `wrist`
camera configuration. The task instruction is normalized to
`push the green button`.

This is the AutoRLBench home of the dataset previously published as
[`villekuosmanen/eval_push_green_button_merged`](https://huggingface.co/datasets/villekuosmanen/eval_push_green_button_merged)
(198 episodes from 15 sources, August 25–31). This revision adds 23 later
eval and RLT green-button runs (1757 episodes).

## Visualization

[Open episode 0 in the LeRobot Dataset Visualizer](https://huggingface.co/spaces/lerobot/visualize_dataset?dataset=AutoRLBench%2Feval_push_green_button_merged&episode=0).

## Success labels

Every source episode has complete per-frame BusyBox state. The episode-level
`success` column is true exactly when `busybox.button.green_button` contains
one rising edge during the episode. Other press counts (0 or ≥2) are labeled
failure. RoboCandyWrapper `busybox_state` sidecars are retained and
renumbered to the merged episode indices.

| policy_type | episodes | successes | success rate |
| --- | ---: | ---: | ---: |
| eval | 332 | 117 | 35.2% |
| rlt | 1623 | 879 | 54.2% |
| **total** | **1955** | **996** | **50.9%** |

## Provenance

`meta/episodes` records `source_repo_id`, `source_revision`,
`source_episode_index`, `policy_type`, `success`, `success_category`,
`green_button_press_count`, `success_source`, `task_id`, and
`task_instruction`.

## Sources

Original 15 sources (August 25–31) plus later `villekuosmanen/eval_push_green_button_*`
and `villekuosmanen/eval_rlt_push_green_button_*` runs with complete BusyBox
coverage and a green-button instruction. One RLT run used tagged language
(`push the green button [base]` / `[frozen_rlt]`); those tags were collapsed
to the canonical instruction. Empty, incomplete, and non-green-button
language runs were left out.