The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
SuperMimic manipulation reset banks
Physically validated start states for three simulated ARX5 manipulation tasks from supermimic_manipulation. Episodes start from these states, following the reset distributions of OmniReset (ICLR 2026).
| File | Split | States |
|---|---|---|
<task>_train.pt |
training | 10,000 per reset category |
<task>_test.pt |
held-out test (different seed and split) | 256 per reset category |
Tasks: lemon, box_in_bin, mug_hang. Each bank has four reset categories: the full task
(object on the table, gripper free), the object grasped on the table, the object grasped in
the air, and the object grasped near the goal (for MugHang, the handle already threaded on
the bar). Each state holds the robot joint positions and velocities, the object pose and
velocity, and the target pose; the bank also records its collection provenance (controller,
physics step and seed). Training refuses a bank whose controller or timing differs from the
current task.
Usage
hf download emopointer/supermimic-manipulation-reset-banks --repo-type dataset --local-dir banks
clab train supermimic --task box_in_bin --reset-bank banks/box_in_bin_train.pt \
--output logs/supermimic/box_in_bin --cuda-graph-mode plain
The banks can be regenerated with clab resets collect (see the repository README).
- Downloads last month
- 30