WuChao-Cauchy/bigym_plus

BiGym Plus Dataset

This dataset is derived from the official BiGym 0.9.0 human demonstrations, covering 40 bimanual humanoid manipulation tasks: 1868 trajectories, 620611 frames, about 25 GB, and roughly 8.6 hours of trajectory at 20Hz.

• Head camera D455 (480×848), fovy: 58, control frequency 20Hz
• Left/right wrist cameras D405 (480×640), fovy: 58, control frequency 20Hz
• state and action in the dataset are uniformly 16-dimensional, isomorphic and co-ordered: floating base (X, Y, Z, RZ; 4) + dual-arm joints (10) + dual grippers (2).
• LeRobot Dataset v3.0 (codebase_version=v3.0).

Task Descriptions

Group 1 · Simple Tasks

Dataset	Task Description	Image	Trajectories	Total Duration (s)
reach_target	Reach the target with either left or right wrist.		60	193.3
reach_target_single	Reach the target with specific wrist.		30	93.2
reach_target_dual	Reach 2 targets, one with each arm.		50	175.7

Group 2 · Long-Horizon Tasks

Multi-stage tasks: pick items from the dishwasher → place them into a cabinet or drawer → close the dishwasher and cabinet doors.

Dataset	Task Description	Image	Trajectories	Total Duration (s)
dishwasher_unload_cups_long	Unload cup from dishwasher in wall cabinet task.		51	1456.5
dishwasher_unload_plates_long	Unload plate from dishwasher in wall cabinet task.		37	1514.0
dishwasher_unload_cutlery_long	Unload cutlery from dishwasher to drawer task.		32	1062.1

Group 3 · Dishwasher Workbench

Centered on the dishwasher: opening/closing its door and trays, and loading/unloading cups, plates and cutlery.

Dataset	Task Description	Image	Trajectories	Total Duration (s)
dishwasher_open	Open the dishwasher door and pull out all trays.		56	605.6
dishwasher_close	Push back all trays and close the door of the dishwasher.		69	586.7
dishwasher_open_trays	Pull the dishwasher's trays with the door initially open.		60	811.4
dishwasher_close_trays	Push the dishwasher's trays back with the door initially open.		60	759.5
dishwasher_load_cups	Load cups to dishwasher task.		60	564.2
dishwasher_unload_cups	Unload cups from dishwasher task.		57	862.9
dishwasher_load_plates	Load plates to dishwasher task.		34	752.6
dishwasher_unload_plates	Unload plates from dishwasher task.		35	1149.4
dishwasher_load_cutlery	Load cutlery to dishwasher task.		34	590.1
dishwasher_unload_cutlery	Unload cutlery from dishwasher task.		39	926.2

Group 4 · Tabletop Manipulation

Opening/closing the drawers and doors of kitchen cabinets, and picking/placing items between cabinets and the countertop.

Dataset	Task Description	Image	Trajectories	Total Duration (s)
drawer_top_open	Open top drawer of the cupboard task.		47	328.2
drawer_top_close	Close top drawer of the cupboard task.		51	168.9
drawers_all_open	Open all drawers of the cupboard task.		54	1053.5
drawers_all_close	Close all drawers of the cupboard task.		60	398.9
wall_cupboard_open	Open doors of the wall cupboard task.		51	382.8
wall_cupboard_close	Close doors of the wall cupboard task.		60	239.4
cupboards_open_all	Open all doors/drawers of the kitchen counter task.		46	1710.7
cupboards_close_all	Close all doors/drawers of the kitchen counter task.		58	1427.0
put_cups	Put cups in the wall cabinet.		56	755.0
take_cups	Take cups from the wall cupboard.		36	571.7
store_box	Put box in the cupboard task.		39	947.0
pick_box	Pick up box from and place it on the counter task.		35	644.7
saucepan_to_hob	Take saucepan from cabinet and place it to hob.		36	688.0
store_kitchenware	Put all kitchenware to cupboard.		33	1061.6
groceries_store_lower	Put groceries to lower cabinets tasks.		32	1606.0
groceries_store_upper	Put groceries to upper cabinets tasks.		35	1090.0
move_plate	Move one plate from one rack to another.		60	216.5
move_two_plates	Move two plates from one rack to another.		51	385.5
stack_blocks	Stack blocks in the correct area of the table.		35	1741.9
flip_cup	Flip cup upside-up task.		60	476.5
flip_cutlery	Flip cutlery item task.		47	544.8
flip_sandwich	Flip sandwich using spatula.		51	1039.6
toast_sandwich	Move sandwich on the frying pan.		35	751.4
remove_sandwich	Remove sandwich from the frying pan.		36	739.5

Group 5 · Combined Tasks

The previous four groups list individual tasks, each with its own dataset directory and a preview image. This group is the four merged dataset directories: the single tasks within each group are merged into one LeRobot dataset, and each trajectory is distinguished by its own English language prompt, so there is no separate task preview image. The directory names match the folders under data/bigym_plus, and the task descriptions are taken verbatim from the prompts actually stored in the dataset's meta/tasks.parquet, ordered by the task_index assigned to the group's sub-tasks at generation time.

Dataset	Task Description	Image	Trajectories	Total Duration (s)
reach	Reach the target with either left or right wrist. Reach the target with specific wrist. Reach 2 targets, one with each arm.	None	140	463.2
long_horizon	Unload cup from dishwasher in wall cabinet task. Unload plate from dishwasher in wall cabinet task. Unload cutlery from dishwasher to drawer task.	None	120	4016.7
dishwasher	Open the dishwasher door and pull out all trays. Push back all trays and close the door of the dishwasher. Pull the dishwasher's trays with the door initially open. Push the dishwasher's trays back with the door initially open. Load cups to dishwasher task. Unload cups from dishwasher task. Load plates to dishwasher task. Unload plates from dishwasher task. Load cutlery to dishwasher task. Unload cutlery from dishwasher task.	None	504	7596.6
tabletop	Open top drawer of the cupboard task. Close top drawer of the cupboard task. Open all drawers of the cupboard task. Close all drawers of the cupboard task. Open doors of the wall cupboard task. Close doors of the wall cupboard task. Open all doors/drawers of the kitchen counter task. Close all doors/drawers of the kitchen counter task. Put cups in the wall cabinet. Take cups from the wall cupboard. Put box in the cupboard task. Pick up box from and place it on the counter task. Take saucepan from cabinet and place it to hob. Put all kitchenware to cupboard. Put groceries to lower cabinets tasks. Put groceries to upper cabinets tasks. Move one plate from one rack to another. Move two plates from one rack to another. Flip cup upside-up task. Flip cutlery item task. Flip sandwich using spatula. Move sandwich on the frying pan. Remove sandwich from the frying pan. Stack blocks in the correct area of the table.	None	1104	18954.1

Generation Method

The original demonstrations were recorded at 500Hz from BiGym's official VR collection (mojo physics simulation, demonstrations.zip v0.9.0). During replay-based generation, the built-in bigym DemoConverter.decimate downsamples the actions from 500Hz to 20Hz. Lightweight demos store only actions and no pixels, so at a 20Hz control frequency each trajectory is replayed with env.reset(seed)+env.step(action) using the seed carried by the demo, rendering frame by frame to capture the three camera images and ensuring reproducibility; the last frame of each trajectory is dropped (no subsequent observation). state/action are uniformly padded to 16 dimensions before being written to the dataset, while env.step still feeds actions of the original dimension (15 or 16) to match the task environment. Replay and evaluation share env_utils.get_state and dim_utils, keeping the training/inference state/action aligned.

Format

• RGB is stored as mp4, av1 codec (libsvtav1 encoder), yuv420p, software encode/decode, no GPU acceleration required.
• Each dataset (single task or combined) has its own directory: meta/{info.json,stats.json,tasks.parquet,episodes/} + data/chunk-{episode_chunk:03d}/file-{file_index:03d}.parquet + videos/{observation.images.cam_high,cam_left_wrist,cam_right_wrist}/chunk-/file-.mp4. info.json records fps, features, total_episodes, total_frames, etc.; tasks.parquet stores task_index→language instruction; stats.json is plain JSON with per-dimension min/max/mean/std and q01/q10/q50/q90/q99.

Dimension Padding

The original tasks fall into two categories: 5 3-DOF tasks (ReachTarget, ReachTargetSingle, ReachTargetDual, MovePlate, MoveTwoPlates, whose base has no Z, 15 dims in total) and 35 4-DOF tasks (16 dims). The 3-DOF tasks are padded at base index 2 with a fixed Z=0.0 (both state and action are padded) to align to 16 dimensions, while 4-DOF tasks are left unchanged; padding applies only to the state/action written to the dataset, and env.step still uses the original dimension. The conversion logic lives in dim_utils.py and is shared between data generation and evaluation.

Upload to HuggingFace Hub

Upload the whole directory as a single dataset repo (replace YOUR_HF_USERNAME with your username):

HF_TOKEN=hf_xxxxx huggingface-cli upload YOUR_HF_USERNAME/bigym_plus \
    /home/chao01.wu/BiGym_Dev/data/bigym_plus --repo-type=dataset

Create the dataset repo on the Hub first, or pass --create-pr to auto-create it.

Download from HuggingFace Hub

--local-dir copies real files to disk (no symlinks) and resumes by default — if interrupted, re-running the same command continues where it left off, without restarting from scratch.

Full download:

HF_TOKEN=hf_xxxxx huggingface-cli download YOUR_HF_USERNAME/bigym_plus \
    --repo-type=dataset --local-dir ./bigym_plus --max-workers 8

Download only specific tasks (glob matches directory names):

HF_TOKEN=hf_xxxxx huggingface-cli download YOUR_HF_USERNAME/bigym_plus \
    --repo-type=dataset --local-dir ./bigym_plus \
    --include "reach_target/*" "reach_target_dual/*"

Common options:

• --max-workers N: concurrent download threads, default 8, can be raised on a good network.
• --include "glob1" "glob2": download only matching files (e.g. "<task>/*" for a single task).
• --exclude "glob1": skip matching files.
• --force-download: re-download even if files already exist locally.
• --revision <commit>: pin a specific commit / branch / tag.
• --quiet: disable progress bars, print only the final path.
• --cache-dir PATH: custom HF cache dir (default ~/.cache/huggingface).

After download, each task subdirectory is a complete lerobot v3.0 dataset, loadable via LeRobotDataset(repo_id="bigym_plus/<task>", root="./bigym_plus").