RoboCasa365 "kitchen8" (openpi format)
440 real human demonstrations from the RoboCasa365 target split, picked for an 8-client federated-learning study. Each client has its own fixed kitchen (layout k, style k; no two clients share a kitchen) and one atomic task with one exact instruction, 55 demos each.
| client | kitchen | skill | task | instruction | demos (of available) | frames |
|---|---|---|---|---|---|---|
| c0 | K2 | Close Door | CloseFridge | "Close the fridge door." | 55 (99) | 12474 |
| c1 | K4 | Turn Lever | TurnOnSinkFaucet | "Turn on the sink faucet." | 55 (70) | 11975 |
| c2 | K8 | Press Button | TurnOnMicrowave | "Press the start button on the microwave." | 55 (65) | 8388 |
| c3 | K9 | Pick & Place | PickPlaceToasterToCounter | "Place the toasted item on a plate." | 55 (90) | 15349 |
| c4 | K1 | Insertion | CoffeeSetupMug | "Pick the mug from the counter and place it under the coffee machine dispenser." | 55 (62) | 16165 |
| c5 | K5 | Close & Open Lid | CloseBlenderLid | "Close the lid blender by securely placing the lid on top." | 55 (74) | 20337 |
| c6 | K7 | Close Door | CloseToasterOvenDoor | "Close the toaster oven door." | 55 (61) | 8988 |
| c7 | K10 | Close & Open Lid | OpenStandMixerHead | "Open the stand mixer head." | 55 (58) | 6946 |
- Source:
ember-lab-berkeley/robocasa365-target-atomic(MIT, the RoboCasa team). Episodes are copied unchanged; the kitchen of each episode comes from RoboCasa's nativeep_meta. Per-client source episode indices:meta/decent_vla_clients.json. - Format: LeRobot v2.1 as written by openpi's pinned lerobot (keys
image=robot0_agentview_left,wrist_image=robot0_eye_in_hand,right_image=robot0_agentview_right, 256x256;state16-D;actions12-D;task= the instruction). Client i owns task index i. - Norm stats:
assets/robocasa365_kitchen8/norm_stats.json, openpi format, taken from the release-wide stats of the full target-atomic set (9,126 episodes). Stats recomputed over this subset would make the rarely-nonzero base-motion action dims degenerate (q01 == q99) and blow up openpi's quantile normalisation. - Made by:
scripts/robocasa365_to_openpi.pyin decent-vla (commit837f31b, branchfeat/robocasa365-openpi-kitchen8); the pick and the arms:scripts/robocasa365_make_kitchen8.py, runnerscripts/robocasa365_kitchen8_h200.sh.
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset # lerobot 0.1.0 (openpi's pin)
ds = LeRobotDataset("DecentVLA/robocasa365_kitchen8_openpi")
Please also cite RoboCasa (Nasiriany et al., "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", RSS 2024) when using this data.
- Downloads last month
- 257