RSC-MyPolicy-checkpoints
All checkpoints and model weights of our own that our RoboSynChallenge submission (liyifreddy/RoboSynChallenge, adapter policy/MyPolicy) uses, one directory per task. The adapter loads them at a pinned revision of this repository (see task_policies, ofi_backbone and ofi_sl_estimator in policy/MyPolicy/deploy_policy.yml); every file is checked against its sha256 before use.
| Task | Policy | Checkpoint |
|---|---|---|
click_bell |
OfI | click_bell/ (this repository) |
drawer_open_place |
SmolVLA | organizers' checkpoint RoboSynChallenge/SmolVLA_sim_drawer_open_place @ c0088d84a568f93fb4401aabafcc41cf643efcdd (not copied here) |
handle_basket |
OfI | handle_basket_mix34/ (since 2026-10-11: the handle_basket/ checkpoint fine-tuned for 34,000 more steps on its 610 demonstrations plus 240 recorded with varied timing); handle_basket/ is the earlier one |
item_assembly |
OfI | item_assembly/ |
items_handover |
OfI | items_handover/ |
manipulate_pipette |
OfI | manipulate_pipette/ |
mixer_operating |
OfI | mixer_operating/ |
sample_loading |
OfI | sample_loading/ (+ sample_loading/rack_estimator/ for its pre-place correction) |
table_rearrangement |
OfI | table_rearrangement/ |
water_pouring |
OfI | water_pouring/ |
Shared by the nine OfI tasks: shared/vjepa2_1_vitl_384_ema_encoder.pt, the EMA encoder weights of the V-JEPA 2.1 ViT-L/16 384 checkpoint by Meta, extracted with values unchanged (V-JEPA 2 is released under the MIT License; its text is in shared/LICENSE-vjepa2).
Each task directory holds one One from Infinity (OfI) policy: a RoboActualizer flow-matching policy (DiT-S head on the frozen V-JEPA 2.1 encoder) fine-tuned from the RoboActualizer RoboTwin checkpoint (db12312607/RoboActualizer, revision fa496fa, MIT) for 160,000 steps (global batch size 32) on demonstrations we generated in the RoboSynChallenge simulator.
File in <task>/ |
What |
|---|---|
step_160000.pt |
trainable weights (the frozen encoder is not included) |
config.yaml |
training configuration (used to build the model; the paths in it are those of the training machine and are not read at inference) |
dataset_stats.json |
action / state normalization statistics |
text_cache/*.pt |
umT5 embedding of the task instruction (Wan2.2 text encoder, Apache 2.0) |
meta/tasks.jsonl |
the task instruction |
ofi_checkpoint.json |
file list with sha256, checked by the adapter before loading |
sample_loading/rack_estimator/: a small rack pose estimator (ResNet-18 + transformer; best.pt weights, wrist_mounts.json fixed wrist camera mounts) that reads the step-0 camera images and the robot's joint and link poses; used only by the optional sample_loading pre-place correction. License of our weights: MIT.