---
license: cc-by-4.0
task_categories:
  - robotics
tags:
  - lerobot
  - robotics
  - imitation-learning
  - vision-language-action
  - amd-rocm
  - genesis
---

# RadeonVLA-Reflex Physical-1K Dataset Card

## Overview

- Dataset name: RadeonVLA-Reflex Physical-1K
- Version: physical-1k-1000
- Generator commit: `dbc12b69e88be787b679bb41cf2026ca43e0eda6`
- Genesis version: 1.1.2
- LeRobot version: 0.6.0
- License: CC BY 4.0
- Public URL: https://huggingface.co/datasets/a3124371940/radeonvla_reflex_physical_1k
- SHA256 (manifest+meta): `5a801c07ac0cee40a9bbd2ee0d5598a8c94a89c3b95e6ec68ab753b5c70246ed`

## Task coverage

The primary L1 dataset has 20 variations: five fruits × four bowl positions.

| Fruit | Bowl positions | Training target | Validation | Held-out evaluation |
|---|---|---:|---:|---:|
| banana | white-left, blue-left, white-right, blue-right | 50 each | 5 each | 10 each |
| lemon | white-left, blue-left, white-right, blue-right | 50 each | 5 each | 10 each |
| plum | white-left, blue-left, white-right, blue-right | 50 each | 5 each | 10 each |
| apple | white-left, blue-left, white-right, blue-right | 50 each | 5 each | 10 each |
| orange | white-left, blue-left, white-right, blue-right | 50 each | 5 each | 10 each |

The release table will replace planned counts with the immutable dataset manifest. L2–L4
data are reported separately and are not implied by the primary L1 total.

## Frame schema

| Field | Shape/type | Description |
|---|---|---|
| observation.images.world | (240, 320, 3) uint8 | World RGB camera |
| observation.images.wrist | (240, 320, 3) uint8 | Wrist RGB camera |
| observation.state | (9,) float32 | Ordered robot and gripper qpos |
| action | (9,) float32 | Absolute joint-position target |
| task | string | Natural-language instruction |
| timestamp | float | Episode time (dataset metadata) |
| episode_index | integer | Episode identifier |
| frame_index | integer | Frame within episode |
| seed | integer | Reset seed in the external per-episode certificate |
| success | boolean | Strict judgement in the external per-episode certificate |

`seed` and `success` are not tensor columns in the LeRobot frame schema. They live in
`certificates/episode_XXXXXX.json`, one atomic certificate for each committed episode.

## Action protocol

- Action type: absolute_joint_position
- Joint names and order: panda_joint1..7, panda_finger_joint1..2
- Dimension: 9
- Joint unit: radians (arm), meters (fingers)
- Gripper convention and range: [0.0, 0.04], open=0.04, closed=0.0
- Control frequency: 20 Hz dataset (sim 100 Hz, decimated)

## Data generation

I collect data with the scripted multi-goal expert (`python -m radeonvla.record_dataset`):

1. expert states follow resolved L1–L4 goals after scene randomization;
2. success requires every fruit center to finish inside the inner bowl footprint after
   at least 60 simulation settle steps;
3. pose jitter is non-overlapping; optional appearance/physics DR flags are supported;
4. recording rate is 20 Hz (sim 100 Hz, decimated);
5. formal collection disables kinematic attachment, placement nudges, and off-table
   respawns; a context guard aborts any rigid-body pose write during the episode;
6. failed episodes are discarded and are not part of Physical-1K;
7. `validate_dataset` checks schema, non-finite values, image statistics, exact 20×50
   coverage, unique seeds, zero interventions, and certificate/episode correspondence;
8. I spot-check camera videos under `datasets/*/videos/` before training;
9. Recording happens under `.inprogress`; the target path is replaced only after finalize,
   coverage checks, and a successful LeRobot reopen.
10. `--resume-incomplete` reconstructs saved counts from LeRobot metadata and reconciles
    the two-phase episode certificates before appending with a fresh seed.

## Split policy

- Strict smoke seeds: 12000–12999
- Training seeds: 20000–29999
- Validation seeds: 40000–40999
- Formal evaluation seeds: 50000–59999
- Interruption/recovery seeds: 60000–60999

No seed may occur in more than one split.

## Quality checks

- no black or corrupt images;
- state and action match the frozen schema;
- no NaN or infinity;
- task, object, and container agree;
- episode success is independently verified;
- random replay videos were manually inspected.

## Assets and limitations

Robot and YCB meshes are populated via `setup_assets` (see `assets/README.md` and
`THIRD_PARTY_NOTICES.md`). The YCB data portal publishes the object models under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/); this release preserves that
license, credits the YCB authors, and notes that blue bowl appearance is applied at scene
build time. The Franka MJCF bundled by Genesis carries Apache-2.0 terms.

This dataset is simulation-only; object and language coverage are limited to the
registered fruit/bowl suite. Final episode/frame counts and immutable revision replace
the measured episode/frame fields after the 1,000-episode validator passed.

## YCB attribution

Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and
Aaron M. Dollar, “The YCB Object and Model Set: Towards Common Benchmarks for
Manipulation Research,” ICAR 2015. Source: https://www.ycbbenchmarks.com/ and
https://ycb-benchmarks.s3-website-us-east-1.amazonaws.com/.