File size: 2,716 Bytes
fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 fe3bb9e f0332a2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | ---
license: mit
pretty_name: RoboTwin ICIL Aloha Clean
task_categories:
- robotics
tags:
- robotics
- imitation-learning
- in-context-learning
- robotwin
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: episodes.jsonl
---
# RoboTwin ICIL Aloha Clean
Training data for [RoboTwin-ICIL](https://github.com/robotensor/RoboTwin-ICIL), a
one-demonstration in-context imitation learning benchmark, in robotwin-icil-hdf5-multiview-v1 format.
## Contents
50 tasks x 50 trajectories = 2,500 aloha-agilex trajectories, all of them training data. Each
episode has an HDF5 trajectory, three aligned 320x240 H.264 videos (head, left wrist, right wrist)
and an instruction JSON. The HDF5 holds joint commands, actions, end-effector poses, camera
calibration and a third-view camera. The three public RGB streams are the policy camera contract.
The dataset viewer rows are `episodes.jsonl`, one per trajectory, pointing at its HDF5 and videos.
There is no validation or test split: policies are scored in simulation, not on stored
trajectories. Evaluation scenes use scene seeds from 100000 upwards, which none of these
trajectories use.
## Representation
See `schema.json`, `manifest.json`, `episodes.jsonl` and `checksums.json`.
Joint state/action is float32 width 14: left six joints, left gripper, right six joints,
right gripper. State is drive targets/gripper commands, not measured encoder positions.
Actions are absolute next-row targets, already shifted. EE pose is width 16, with xyz,
quaternion wxyz and gripper for each arm. Decode the HDF5 JPEGs as standard RGB, without channel
reversal.
Videos have exactly N frames aligned to the N HDF5 transition rows. They are lossy convenience
views; HDF5 JPEGs remain authoritative training RGB. Their 30 fps is presentation timing only.
Simulation timestamps are unavailable, and the stored frequency value is not verified physical Hz.
No terminal wrist frame or initial scene fingerprint exists.
## Evaluation
`robotwin-icil standard-eval` scores a frozen policy on all 50 tasks, 100 episodes each, in a clean
and a randomized setting. Each episode generates an expert demonstration in simulation, rebuilds
the identical scene and gives the policy that demonstration as context. Stored trajectories are
not usable as evaluation `prompt.npz` files. Task/seed/instruction/calibration/scene metadata must
not reach the policy. Use the neutral instruction: "Follow the demonstrated behavior."
## License
MIT; see `LICENSE`, which carries the copyright notice of RoboTwin 2.0, the simulation platform
these trajectories come from. `provenance.json` records their origin. Simulator and object assets
are not included.
|