twobidoubi/Robo_model

RoboSynChallenge team checkpoints (π0.5)

π0.5 checkpoints fine-tuned by our team for the RoboSynChallenge CobotMagic simulation tasks, trained with openpi from pi05_base on the official RoboSynChallenge/cobotmagic_Sim_<task> datasets (1,000 expert episodes per task, LeRobot v2.1). The repository has one folder per task (ten tasks); inside, <run>/<step>/ is one openpi checkpoint step: params/ (full model parameters, also for LoRA runs), assets/<repo_id>/norm_stats.json and, for most folders, train_state/ (optimizer state) for resuming.

Checkpoints (<task>/<run>/<step>)

Path	Task	Run	Step	Success / episodes	Mean action steps / limit	Evaluation
table_rearrangement/pi05_full_v1/5000	table_rearrangement	full fine-tune	5000	60 / 100	246.6 / 361	random setting, 100 episodes
table_rearrangement/pi05_full_v1/10000	table_rearrangement	full fine-tune	10000	73 / 100	219.1 / 361	random setting, 100 episodes
table_rearrangement/pi05_full_v1/15000	table_rearrangement	full fine-tune	15000	78 / 100	209.3 / 361	random setting, 100 episodes
table_rearrangement/pi05_full_v1/19999	table_rearrangement	full fine-tune	19999	76 / 100	229.3 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/2500	table_rearrangement	LoRA	2500	32 / 100	299.2 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/5000	table_rearrangement	LoRA	5000	39 / 100	294.6 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/7500	table_rearrangement	LoRA	7500	52 / 100	259.5 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/10000	table_rearrangement	LoRA	10000	56 / 100	265.4 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/12500	table_rearrangement	LoRA	12500	70 / 100	233.6 / 361	random setting, 100 episodes
table_rearrangement/pi05_lora_v1/14999	table_rearrangement	LoRA	14999	76 / 100	223.9 / 361	random setting, 100 episodes
items_handover/pi05_lora_v1/2500	items_handover	LoRA	2500	7 / 100	348.3 / 350	random setting, 100 episodes
items_handover/pi05_lora_v1/5000	items_handover	LoRA	5000	1 / 100	349.5 / 350	random setting, 100 episodes
items_handover/pi05_lora_v1/7500	items_handover	LoRA	7500	11 / 100	349.0 / 350	random setting, 100 episodes
items_handover/pi05_lora_v1/10000	items_handover	LoRA	10000	14 / 100	347.6 / 350	random setting, 100 episodes
items_handover/pi05_lora_v1/12500	items_handover	LoRA	12500	38 / 100	342.2 / 350	random setting, 100 episodes
items_handover/pi05_lora_v1/14999	items_handover	LoRA	14999	14 / 100	348.4 / 350	random setting, 100 episodes
items_handover/pi05_aug_lora_v1/2500	items_handover	LoRA (team-generated data)	2500	0 / 100	350.0 / 350	random setting, 100 episodes; trained on the team-generated data (see Notes)
items_handover/pi05_aug_lora_v1/5000	items_handover	LoRA (team-generated data)	5000	2 / 42	—	early-stopped
items_handover/pi05_aug_lora_v1/7500	items_handover	LoRA (team-generated data)	7500	0 / 25	—	early-stopped
items_handover/pi05_aug_lora_v1/10000	items_handover	LoRA (team-generated data)	10000	0 / 25	—	early-stopped
items_handover/pi05_aug_lora_v1/12500	items_handover	LoRA (team-generated data)	12500	1 / 29	—	early-stopped
items_handover/pi05_aug_lora_v1/14999	items_handover	LoRA (team-generated data)	14999	0 / 25	—	early-stopped
handle_basket/pi05_lora_v1/2500	handle_basket	LoRA	2500	28 / 100	472.9 / 500	random setting, 100 episodes
handle_basket/pi05_lora_v1/5000	handle_basket	LoRA	5000	35 / 100	467.1 / 500	random setting, 100 episodes
handle_basket/pi05_lora_v1/7500	handle_basket	LoRA	7500	79 / 100	414.0 / 500	random setting, 100 episodes
handle_basket/pi05_lora_v1/10000	handle_basket	LoRA	10000	45 / 100	447.2 / 500	random setting, 100 episodes
handle_basket/pi05_lora_v1/12500	handle_basket	LoRA	12500	25 / 100	469.4 / 500	random setting, 100 episodes
handle_basket/pi05_lora_v1/14999	handle_basket	LoRA	14999	65 / 100	431.6 / 500	random setting, 100 episodes
item_assembly/pi05_lora_v1/2500	item_assembly	LoRA	2500	0 / 18	— / 500	early-stopped at 0/18 (no realistic chance against the baselines), gripper rescaling on
item_assembly/pi05_lora_v1/5000	item_assembly	LoRA	5000	0 / 18	— / 500	early-stopped at 0/18 (no realistic chance against the baselines), gripper rescaling on
item_assembly/pi05_lora_v1/7500	item_assembly	LoRA	7500	0 / 25	— / 500	early-stopped at 0/25 (no realistic chance against the baselines), gripper rescaling on
item_assembly/pi05_lora_v1/10000	item_assembly	LoRA	10000	16 / 100	469.1 / 500	random setting, 100 episodes, gripper rescaling on
item_assembly/pi05_lora_v1/12500	item_assembly	LoRA	12500	2 / 40	— / 500	early-stopped at 2/40 (no realistic chance against the baselines), gripper rescaling on
item_assembly/pi05_lora_v1/14999	item_assembly	LoRA	14999	13 / 100	474.6 / 500	random setting, 100 episodes, gripper rescaling on

Recommended checkpoint per task

Task	Best of ours	Official / community baselines (per 100)
table_rearrangement	table_rearrangement/pi05_full_v1/15000 (78/100)	ACT 96, SmolVLA 90, DP 17
items_handover	items_handover/pi05_lora_v1/12500 (38/100)	ACT 26, SmolVLA 22, DP 0
handle_basket	handle_basket/pi05_lora_v1/7500 (79/100)	ACT 54, SmolVLA 33, DP 12
item_assembly	item_assembly/pi05_lora_v1/10000 (16/100, below every baseline)	ACT 65, community π0.5 64, SmolVLA 51, DP 38

Training recipes

Run	Trainable weights	Steps	Batch	LR schedule	EMA	Wall clock	Checkpoint size
table_rearrangement/pi05_full_v1 (full)	all 3.35B parameters	20,000	64	cosine 2.5e-5 → 2.5e-6 over 30k, warmup 1k	0.99	34 h 58 min, one shared H100 NVL	42 GB (params 12 GB + train_state 31 GB); the step-5000 folder holds params only
<task>/pi05_lora_v1 (LoRA)	LoRA rank 16 (PaliGemma) / 32 (action expert) + SigLIP encoder + projections, 467M of 3.4B	15,000 (last save named 14999)	64	cosine 2.5e-5 → 2.5e-6 over 15k, warmup 1k	off	8 to 13 h on two H100 NVL	9.5 GB (params + optimizer state of the trained subset)

Action horizon 50, images 224×224 from three cameras, 14-D state and absolute joint actions with the openpi delta transform for the arm joints (config names pi05_<task> and pi05_<task>_lora in policy/pi05/src/openpi/training/config.py of the team repository).

Using 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/pi05_lora_v1/7500      # -> checkpoints/handle_basket/pi05_lora_v1/7500
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/pi05_lora_v1/7500 --config-name pi05_handle_basket_lora --label hb_7500 --rescale-gripper

Pass the openpi config the checkpoint was trained with (--config-name pi05_<task> for the full fine-tune, pi05_<task>_lora for LoRA). The worker executes 10 of the 50 predicted actions per call.

Gripper scaling. The official datasets store the two gripper targets in [0, 1] for Items hand-over, Handle basket and Item assembly, while the simulator clips actions to [0, 0.05]; pass --rescale-gripper for those tasks. Table rearrangement stores physical gripper positions and must not use it. (The Items hand-over and Handle basket numbers above were measured without rescaling, which happens to act as an open/close threshold; Item assembly needs the rescaling.)

Notes

• Success rates come from our own 100-episode evaluations in the random setting with expert feasibility screening, on the same EmbodiChain simulator as the official baselines; they are not official scores. A 100-episode run has roughly a 5-point standard error.

• LoRA success is not monotonic in training steps on these tasks (e.g. Handle basket 79 at step 7500, then 45, 25, 65) while the training loss falls smoothly; pick checkpoints by evaluation, not by loss.

• Replays, training records, loss curves and per-checkpoint analyses are on the team page (task pages, "Policies" and "Our fine-tuning runs").

• Datasets for augmentation experiments are in 2bidoubi/Robo_trajectory; this repository holds model weights only.

• items_handover/pi05_aug_lora_v1/* was trained with the same LoRA recipe on the team-generated robotrajectory/RoboSyn-Trajectory folder items_handover_augmented_1702 (1702 episodes) instead of the official release. Those episodes have no appearance randomisation (white table, grey holder, black pen, fixed layout, no distractors), while the official data and the evaluation randomise textures, colours, poses and distractors, so every checkpoint scored 0 to 2 successes; evaluations after step 2500 were stopped early (0 successes after 25 episodes or at most 2 after 40). Kept for the record only.