---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- lerobot
- v3.0
- robotics
- ur7e
- universal-robots
- real-robot
- robotiq-2f85
- realsense
- multi-camera
- code-as-policies
- vla
- smolvla
- imitation-learning
- manipulation
- stacking
- absolute-action
- skill-target-action
- corl2026-csi
pretty_name: UR7e CaP Stack Block 100 Episodes
size_categories:
- 100<n<1K
---

# UR7e CaP Stack Block 100 Episodes

This dataset contains 100 clean forward-success demonstrations for a real UR7e stacking task, recorded in LeRobot v3.0 format.

Task instruction: **Build a stack on the blue dish with the red block on the bottom, the green block in the middle, and the blue block on top.**

## Dataset Summary

- Robot: Universal Robots UR7e
- Gripper: Robotiq 2F-85 style gripper signal represented as a ratio
- Cameras: RealSense topview and RealSense wrist, stored as video features
- Format: LeRobot v3.0
- Nominal FPS: 30
- Episodes: 100 selected clean forward episodes
- Action convention: absolute target action, matching `skill.goal_position.joint`; this is not a delta-action dataset
- Task type: real-robot block stacking with partial closed-loop redetection between stack stages

## Features

Core learning fields include:

- `observation.state`: 6 UR7e arm joint positions plus gripper ratio
- `action`: absolute target joint action, 6 arm joint targets plus gripper target
- `observation.images.realsense_topview`: top-view RGB video
- `observation.images.realsense_wrist`: wrist RGB video
- `skill.type`, `skill.progress`, `skill.goal_position.*`, and `skill.natural_language`: skill-level target/context annotations
- `observation.ee_pos.robot_xyzrpy` and `observation.gripper_binary`: auxiliary robot state annotations

## Quality Notes

The uploaded split intentionally excludes early legacy recordings that had irregular frame timing, missing per-frame camera synchronization audit metadata, and compressed one-frame gripper transitions. The selected 100 episodes passed local checks for raw-file completeness, frame/image count consistency, finite state/action values, monotonic timestamps, camera sync metadata presence, and successful forward-task judging.

Partial closed-loop stack stages intentionally pause and redetect between stages. Those detection intervals are not part of the recorded forward motion segments; the uploaded episodes contain the robot execution data intended for VLA tuning.

## Provenance

A source manifest is included at `meta/clean_source_manifest.json`, mapping every uploaded LeRobot episode back to its local raw session and episode number.