ReMemBench-Migrated

Regenerated version of the ReMemBench dataset, replayed in the latest RoboCasa (v1.0.1 / RoboCasa365, robosuite 1.5.2, mujoco 3.3.1) environment. The original ReMemBench demos were recorded in RoboCasa 0.2.0; this dataset replays each demonstration's recorded MuJoCo states through the migrated RoboCasa 1.0.1 memory-task environments and re-derives observations and images.

Contents

13 memory tasks, 25 source files, 421 demonstrations, all successful. Layout mirrors the original: <task>/<date>/{demo.hdf5, demo_im128_notp.hdf5}.

• demo.hdf5 — robomimic low-dim convention (data/demo_<i>/): states, actions, action_dict/, action_infos/, observations/ (low-dim), policy_mode, initial_obj_qpos/; per-demo attrs ep_meta/initial_qpos/model_file/ non_robot_qpos_idx/num_samples.
• demo_im128_notp.hdf5 — robomimic image convention (data/<Env>_<Robot>_demo_<i>/): adds 128×128 RGB from 4 cameras (robot0_agentview_center, robot0_agentview_left, robot0_agentview_right, robot0_eye_in_hand) plus rewards/dones. RGB only (no segmentation / bbox suite).

Tasks: MemFruitInSinkLeftFar, MemFruitInSinkRightFar, MemHeatPot, MemHeatPotMultiple, MemPutKBowlInCabinet, MemPutKBreadInMicrowave, MemRetrieveOilsFromCounter{LL,LR,RL,RR}, MemWashAndReturn{Left,Right,SameLocation}.

Notes

• State-faithful replay: trajectories are the original recorded states; visuals reflect the migrated RoboCasa 1.0.1 env.
• Container-lift correction: containers (plates/bowls) whose recorded z clipped through the thinner 1.0.1 counter slabs are lifted so they rest on the surface; this lift is baked into the saved states so states/obs/images agree. All other qpos are unchanged.
• Some 0.2.0 fixtures fall back to 1.0.1 defaults, so scenes are visually close but not pixel-identical to the originals.

The migrated task environments live in yilin-wu98/robocasa (branch feat/halo-memory-tasks).