BimanualYAM-models (RoboColosseum)

These are vision-language-action (VLA) models fine-tuned on BimanualYAM manipulation tasks for the RoboColosseum benchmark. Each VLA is fine-tuned from its public base checkpoint with its repo-default recipe: default LRs and frozen parts, with the LR scaled linearly with the effective batch size, for 5 epochs. Each folder holds the final-epoch inference weights plus the small config files needed to load the model.

Layout: <model>/<task>/

Model	Base checkpoint	Dustpan	Microwave	Drawer	Cups	Tray
gr00t/ GR00T N1.7	nvidia/GR00T-N1.7-3B	✅	✅	✅	✅	✅
pi05/ π0.5	gs://openpi-assets/checkpoints/pi05_base	✅	✅	✅	✅	✅
g05/ G0.5	OpenGalaxea/G05 (g05-base)	✅	✅	✅	✅	✅
molmoact2/ MolmoAct2	allenai/MolmoAct2	✅	✅	✅	✅	✅
lingbot-vla-v2/ LingBot-VLA v2 6B	robbyant/lingbot-vla-v2-6b	✅	✅	✅	✅	✅

Each <model>/<task>/README.md gives the recipe, the files, and how to load the model.

Tasks (RoboColosseum/BimanualYAM-datasets, LeRobot v3.0, 30 fps)

Task	Instruction	Episodes	Frames
Dustpan	"Clean the table."	101	55,046
Microwave	"Open the microwave and take out the bowl."	100	108,487
Drawer	"Put the cup into the drawer."	101	82,305
Cups	"Stack the cups."	100	39,184
Tray	"Clean the table."	102	50,254

Common setup:

• Cameras: top (360x640), left wrist and right wrist (480x640).
• State and action: 14-D absolute joint angles [left arm 6, left gripper 1, right arm 6, right gripper 1].
• Training: fp32 master weights with bf16 compute, on 4x H100.

Recipes (same for every task)

Model	Trained parts	GBS	Peak LR
GR00T N1.7	Full: VLM backbone unfrozen + action head	64	1e-4, cosine, 5% warmup
π0.5	Full (openpi), EMA 0.99	64	2.5e-5, cosine to 2.5e-6, 1k warmup
G0.5	Full (vision LR x0.1), FSDP	32	4e-5, cosine, 200 warmup, wd 0.03
MolmoAct2	Full	128	LLM 2e-5, ViT 1e-5, connector 1e-5, action expert 1e-4
LingBot-VLA v2	Full	64	1.25e-5 (5e-5 x 64/256), constant, Muon

Training logs are in the W&B project shaileshxml-nus/RoboColosseum.