Datasets:

HBOrtiz
/

so101_eval1

license: mit
task_categories:
  - robotics
tags:
  - lerobot
  - so-101
  - imitation-learning
  - behavior-cloning
  - hg-dagger
size_categories:
  - 10K<n<100K

so101_eval1

Training dataset for Eval 1 of the LeMonkey project (ETH Robot Learning course, FS26): a SmolVLA policy that picks up a banana and places it in the bowl named by the prompt.

153 teleop episodes, 44,604 frames, 13 task strings.
Source robot: SO-101 (6-DOF arm), wrist-mounted USB camera at 480x640 at 30 fps.
Format: LeRobot v3 (codebase_version: v3.0).
Aggregated from 6 per-color recording sessions:
- 118 behavior-cloning demos (blue / red / green, ~39 each)
- 35 HG-DAgger correction demos (blue / red / green) recorded against the behavior-cloning policy's failure positions
The DAgger sources were normalized to the BC schema (rename camera1->front, drop is_intervention) before aggregation.

Trained policy

The deployed Eval 1 policy is HBOrtiz/so101_smolvla_eval1: SmolVLA-450M, 25k steps from lerobot/smolvla_base, image augmentation on.

Prompt templates

Put the banana in the blue colored bowl.
Put the banana in the red colored bowl.
Put the banana in the green colored bowl.
(+ paraphrases and DAgger correction prompts)

Repository

Project page and code: github.com/Ace3Z/LeMonkey