Bluexxx/robocasa365_kitchen18

RoboCasa365 Kitchen-18 (stereo agentview pair)

The Kitchen-18 protocol of the paper StereoLift: Enhancing Vision-Language-Action Models for Precise Manipulation via Calibration-Free Stereo Lifting: the 18 atomic target tasks of the RoboCasa365 Target50 benchmark with 100 human demonstrations per task, drawn from the RoboCasa365 Human300 pretraining set, in LeRobot v3 format with the head-mounted stereo pair robot0_agentview_left / robot0_agentview_right (256 × 256 pixels, 20 Hz).

Code, rectification files and launch scripts: https://github.com/peacelwh/StereoLift (see docs/ROBOCASA365.md).

Contents

Episodes	1,800 (18 tasks × 100 human demonstrations)
Frames	406,831 at 20 Hz
Cameras	observation.images.robot0_agentview_left, observation.images.robot0_agentview_right (256 × 256 × 3, H.264)
State	observation.state, 16-d (base pose, base-relative end-effector pose, gripper)
Action	action, 12-d (base motion, control mode, end-effector delta, gripper)
Language	task (the instruction), annotation.human.task_description, annotation.human.task_name
Robot	PandaOmron (mobile Franka arm with a parallel gripper)

Episodes 100 k .. 100 k + 99 belong to task k of the list below (the order of configs/kitchen18_tasks.json in the code repository):

 0 CloseBlenderLid            6 OpenDrawer                 12 PickPlaceToasterToCounter
 1 CloseFridge                7 OpenStandMixerHead         13 SlideDishwasherRack
 2 CloseToasterOvenDoor       8 PickPlaceCounterToCabinet  14 TurnOffStove
 3 CoffeeSetupMug             9 PickPlaceCounterToStove    15 TurnOnElectricKettle
 4 NavigateKitchen           10 PickPlaceDrawerToCounter   16 TurnOnMicrowave
 5 OpenCabinet               11 PickPlaceSinkToCounter     17 TurnOnSinkFaucet

Extra files:

• meta/task_manifest.parquet: the 18 task blocks, for the RoboCasa365 task-balanced sampler (--dataset.task_sampling_manifest).
• meta/stats_task_balanced.json: task-balanced mean and standard deviation of state and action (every task weighs the same), the normalisation used in the paper (--dataset.stats_override_path).
• kitchen18_30demos.json and meta/stats_task_balanced_30demos.json: the 30-demonstration protocol (the first 30 episodes of every task) and its statistics.

Protocol

Policies are trained jointly on the 18 tasks (global batch 64, cosine learning-rate schedule defined over 20,000 steps) and evaluated with 50 rollouts per task in the official pretrain scene split with the official object registries, seed 7 with per-episode offsets, the official per-task horizon and re-planning every 5 steps of a 50-step action chunk; results are reported as the mean success over the 18 task means with its standard error, the mean over the five pick-and-place tasks and paired per-task differences.

How it was built

Extracted from the LeRobot conversion of Human300 (Whalswp/robocasa_pretrain_human300_v3; official release at ember-lab-berkeley/robocasa365-pretrain-*) with lerobot.datasets.dataset_tools.split_dataset, keeping the episodes 100 i .. 100 i + 99 of the 18 tasks (Human300 orders its 300 tasks alphabetically inside each family, 65 atomic then 235 composite), then dropping the robot0_eye_in_hand camera with remove_feature. Frames, states, actions and instructions are unchanged.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Bluexxx/robocasa365_kitchen18")
print(ds.meta.total_episodes, ds.meta.camera_keys)

License and attribution

The demonstrations are part of RoboCasa365 (https://robocasa.ai) and keep its terms; this redistribution is for non-commercial research. Please cite RoboCasa365 and, if you use this subset or the code, StereoLift:

@misc{li2026stereolift,
  title  = {StereoLift: Enhancing Vision-Language-Action Models for Precise Manipulation via Calibration-Free Stereo Lifting},
  author = {Li, Wenhao and Liang, Xusheng and Shi, Yang and Han, Zhongyi and Qin, Zheyun and Yin, Yilong},
  year   = {2026},
  howpublished = {\url{https://github.com/peacelwh/StereoLift}}
}