twobidoubi/Robo_best_model

RoboSynChallenge team submission: best π0.5 checkpoint per task

One folder per task, each holding the best checkpoint our team has for that task: <task>/params/ (openpi checkpoint, full model parameters, bfloat16) and <task>/assets/<repo_id>/norm_stats.json. No optimizer state. Every checkpoint was fine-tuned with openpi from pi05_base on the official RoboSynChallenge/cobotmagic_Sim_<task> dataset (1,000 expert episodes), evaluated by us for 100 episodes in the random setting on the EmbodiChain simulator. The complete checkpoint series (every saved step, with optimizer state) are in twobidoubi/Robo_model; training records, loss curves and replays are on the team page.

Folder	Task	Recipe	Source (series/step)	Our success / 100	Official baselines (ACT / SmolVLA / DP)	openpi config	--rescale-gripper
table_rearrangement/	Table rearrangement	full fine-tune, 20k steps	table_rearrangement/pi05_full_v1/15000	78	96 / 90 / 17	pi05_table_rearrangement	no
click_bell/	Click bell	LoRA, 15k steps	click_bell/lei_lora_v1/14999	92	74 / 43 / 66	pi05_click_bell_lora	no
handle_basket/	Handle basket	LoRA, 15k steps	handle_basket/pi05_lora_v1/7500	79	54 / 33 / 12	pi05_handle_basket_lora	yes
items_handover/	Items hand-over	full fine-tune, 20k steps	items_handover/pi05_full_v1/15000 (in 2bidoubi/Robo_model)	78	26 / 22 / 0	pi05_items_handover	yes
drawer_open_place/	Drawer open and place	LoRA	drawer_open_place/lei_lora_B_v1/checkpoint_005000	28 / 97 (development seeds)	41 / 37 / 5	pi05_drawer_open_place_lora	yes
item_assembly/	Item assembly	LoRA, 15k steps	item_assembly/pi05_lora_v1/10000	16	65 / 51 / 38	pi05_item_assembly_lora	yes
manipulate_pipette/	Manipulate pipette	LoRA, 15k steps	manipulate_pipette/lei_lora_v1/14999	41	64 / 25 / 46	pi05_manipulate_pipette_lora	yes

Bold: above every official baseline. Water pouring, Mixer operating and Sample loading have no team checkpoint in this repository yet.

Common settings: three RGB cameras at 224×224, 14-D joint state, prompt from the task name, action horizon 50 with 10 actions executed per inference call, arm actions as deltas relative to the current state (restored before execution), gripper actions absolute. --rescale-gripper maps the [0, 1] gripper targets of the dataset to the simulator's [0, 0.05] range for the tasks marked "yes". The lei_* folders also carry an inference_config.json with the same information.

Running a checkpoint

From the team repository (arlo-yang/RoboSynChallenge), with the π0.5 inference worker installed (policy/pi05_community/setup.sh):

python web/scripts/hf_hub.py download checkpoint handle_basket --repo twobidoubi/Robo_best_model   # -> checkpoints/handle_basket
conda activate robosyn && export ROBOSYN_RUNTIME=$CONDA_PREFIX
bash policy/pi05_community/inference.sh eval --task handle_basket --episodes 100 --gpu 0 \
  --local-checkpoint checkpoints/handle_basket --config-name pi05_handle_basket_lora --label best_hb --rescale-gripper

Pass the config name from the table, and --rescale-gripper only where the table says yes. The worker executes 10 of the 50 predicted actions per call.

Notes

• Success rates are our own 100-episode evaluations in the random setting with expert feasibility screening; a 100-episode run has roughly a 5-point standard error. They are not official scores.
• Items hand-over: the score was measured on 2026-10-11 after a broken success check in our local task code was repaired; the full series and the corrected numbers of every checkpoint are in 2bidoubi/Robo_model.
• Drawer open and place: Lei's selected checkpoint; the 28/97 figure comes from 100 development seeds with three environment reset errors excluded.