RoboSynChallenge MIAA-HCC checkpoints
The current MIAA-HCC submission uses ACT for Item Assembly and task-specific pi0.5 policies for the other nine tasks. Click Bell additionally uses a lightweight three-view RGB/proprioception feedback model for bounded correction and pressing. Manipulate Pipette reuses pi0.5 with observation-gated waypoint pressing and model-guided lateral correction. Runtime adapters, locked dependencies, installation steps, task settings, and evaluation commands are provided in RoboSynChallenge_MIAA-HCC, at the beginning of its README.
Current submission mapping
| Task | Policy | HF folder | Checkpoint source |
|---|---|---|---|
| click_bell | pi0.5 + RGB feedback | click_bell/ + click_bell_servo/ |
Original official 20k, source step 19999 |
| items_handover | pi0.5 | items_handover/ |
Original official task-specific 20k, source step 19999 |
| water_pouring | pi0.5 | water_pouring/ |
Original official 20k, source step 19999 |
| handle_basket | pi0.5 | handle_basket/ |
Official data revision 2026-09-04, retrained 20k, source step 19999 |
| mixer_operating | pi0.5 | mixer_operating/ |
A_blank_seed260813, source step 19999 |
| item_assembly | ACT | item_assembly_act/ |
Complete ACT checkpoint; see its folder README |
| table_rearrangement | pi0.5 | table_rearrangement/ |
Original official 20k, source step 19999 |
| drawer_open_place | pi0.5 | drawer_open_place/ |
Original official 20k, source step 19999 |
| manipulate_pipette | pi0.5 + waypoint feedback | manipulate_pipette/ |
Original official 20k, source step 19999; adapter details in GitHub README |
| sample_loading | pi0.5 | sample_loading/ |
Original official 20k, source step 19999 |
The pi0.5 folders preserve params/, assets/, and _CHECKPOINT_METADATA. Place them directly under policy/pi05/checkpoints/pi05_base_robosynchallenge_full/<task>/; no /19999 layer is required.
For ACT, place the contents of item_assembly_act/ directly under policy/act/checkpoints/item_assembly_act/.
Use the matching normalization statistics, camera preprocessing, execution settings, and gripper conversion supplied by the corresponding adapter. Training optimizer states are omitted.
Manipulate Pipette — Model-Guided Waypoint Pressing
Manipulate Pipette automatically uses observation-gated waypoint pressing and
model-guided lateral correction in the current GitHub main adapter. It reuses
the existing manipulate_pipette/ pi0.5 checkpoint. No additional learned model,
weights, or training data is required. During inference, the policy/controller
uses three camera RGB views, the robot's own joint state, and the task instruction.
The frozen setup uses a 50-action prediction horizon, chunk 20, 5 denoising steps, and normalized output grippers. An observed-descent gate can admit one takeover; pressing waypoints use 0.25 mm increments, with refreshed model guidance every 20 actions and lateral corrections limited to 0.05 mm per action and a 2 mm target offset. Motion and IK checks clear stale commands and return to fresh pi0.5 inference.
Installation, locked dependencies, and the evaluation command are at the beginning of the GitHub README.
For reproducibility, continue downloading the tested weight revision
ae4c33026ca1e7e719e158758e9e14a652f991cd.
Click Bell — Visual Feedback Control
We trained a lightweight three-view CNN (1,089,488 parameters) to predict local
position corrections, visibility, and uncertainty. During inference, the
feedback controller uses only the three camera RGB images and the robot's own
14-D joint state. It monitors fresh observations at every step and can take over
from pi0.5 for bounded correction and pressing, then return to pi0.5 when needed.
The Click Bell adapter automatically uses pi0.5 with the frozen
vision_extend feedback controller. Pi0.5 retains
its 50-step prediction horizon and 25-action execution prefix.
Download both click_bell/ (the original pi0.5 checkpoint) and
click_bell_servo/model.pt (the auxiliary inference checkpoint). Place the latter
at policy/pi05/checkpoints/click_bell_servo/model.pt. The official adapter loads
both automatically for Click Bell. The auxiliary checkpoint includes the model,
normalization/configuration, and uncertainty calibration; it omits optimizer and
RNG states. The original Click Bell pi0.5 weights are retained.
The auxiliary model was trained using newly collected samples from the official RoboSynChallenge simulator (200 training scenes and 40 validation scenes, 5,340 three-view samples) and 6,400 auxiliary frames from 800 historical training demonstration episodes.
Installation, the locked pi0.5 environment, action conversion, and the official evaluation command are documented at the beginning of the GitHub README.
Initial repository preparation: 2026-10-06. ACT submission mapping updated: 2026-10-11.
Click Bell RGB feedback integration updated: 2026-10-12.
Manipulate Pipette main integration updated: 2026-10-12.