Dataset Viewer
Auto-converted to Parquet Duplicate
root
string
codebase_version
string
total_episodes_info
int64
total_frames_info
int64
episode_rows
int64
episode_length_sum
int64
dataset_len
int64
data_files
int64
video_files
int64
unique_video_files_by_key
dict
episodes_per_data_file_max
int64
episodes_per_video_file_max_by_key
dict
video_path_template
string
data_path_template
string
samples
list
/home/zrgong/projects/RMBench-Plus-Plus/generated_data_lerobot_v3_put_back_block_relocate_v1/put_back_block_relocate
v3.0
100
47,155
100
47,155
47,155
100
400
{ "observation.images.head_camera": 100, "observation.images.left_camera": 100, "observation.images.right_camera": 100, "observation.images.front_camera": 100 }
1
{ "observation.images.head_camera": 1, "observation.images.left_camera": 1, "observation.images.right_camera": 1, "observation.images.front_camera": 1 }
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
[ { "index": 0, "episode_index": 0, "frame_index": 0, "observation_state_shape": [ 14 ], "action_shape": [ 14 ], "image_shapes": { "observation.images.head_camera": [ 3, 240, 320 ], "observation.images.left_camera": [ 3, ...

RMBench++ Put Back Block Relocate v1

This pilot dataset contains 100 official-RT scripted-expert demonstrations for the frozen put_back_block_relocate training split.

The policy must pick up the block, move it to the center, press a button with the left hand, and remember the brief post-button feedback. If one mat lights up, the block must be placed on that newly indicated mat. If no mat lights up, the block must be returned to the originally indicated mat.

Canonical Task Prompt

Pick up the block, move it to the center, press the button, watch for a brief mat signal, and put the block back on the remembered target mat.

Frozen Coverage And Contents

  • 100 training episodes using the frozen v1 training manifest.
  • Four initial mats are balanced: 25 episodes each.
  • Five relocate outcomes are balanced: left, right, front, back, and no-light occur 20 times each.
  • The effective target mat is balanced: 25 episodes each.
  • One Parquet file per episode and one H.264 MP4 per episode per camera.
  • Native RMBench 14D observation.state and 14D action.
  • Policy RGB views: head, left wrist, and right wrist. The front camera is included only for QA and visualization.
  • Per-frame episode seed, canonical prompt provenance, JSON-encoded rmbench_plus audit metadata, and memory-keyframe sidecars under meta/.

The embedded meta/rmbench_plus_manifest.json contains the frozen train and eval seed specifications. This repository publishes training demonstrations only; eval expert trajectories are not included.

Memory Keyframes

meta/memory_keyframes.json uses schema rmbench_memory_keyframes_v1. Each episode contains task/keyframe labels for discriminator-style FULL-frame retention training:

  • initial_belief
  • center_place
  • button_press_down
  • memory_update_target or memory_confirm_no_update
  • terminal final_place

The companion meta/memory_keyframe_audit.json records role counts and protocol version.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="zrgong/rmbench-plus-plus-put-back-block-relocate",
    video_backend="pyav",
)

Validation And Provenance

Before upload, the production gate required 100/100 training scripted-expert success, 100/100 private eval scripted-expert preflight success, exact manifest metadata for every episode, native 14D reads, one data shard per episode, four video streams per episode, and a complete LeRobot v3 validation summary.

The exact RMBench++ Git revision used for production is stored in meta/rmbench_plus_source_revision.txt.

These demonstrations are training and scripted-expert evidence. They are not closed-loop policy evaluation results, and this v1 pilot is not a general-memory leaderboard.

Downloads last month
-