RoboColosseum/droid-open-drawer-molmoact2

DROID-models-10ep (RoboColosseum)

These are vision-language-action (VLA) models fine-tuned on two DROID (Franka) manipulation tasks for the RoboColosseum benchmark, for 10 epochs in total. Each model is the 5-epoch model from DROID-models, trained for 5 more epochs:

• The recipe, global batch size and learning rates are the same as the 5-epoch run, which uses the same LRs as BimanualYAM-models.
• Training starts from the 5-epoch final weights, with a fresh optimizer and the LR schedule restarted from the beginning.
• Each run used 1 H200. Gradient accumulation keeps the original global batch size.
• Each run logs into the same W&B run as its 5-epoch phase.

Each folder holds the final inference weights plus the small config files needed to load the model.

Layout: <model>/<task>/

Model	Base checkpoint	MugOnCoaster	CloseLaptop	Size per folder
gr00t/ GR00T N1.7	nvidia/GR00T-N1.7-3B	✅	✅	12.6 GB
pi05/ π0.5	gs://openpi-assets/checkpoints/pi05_base	✅	✅	12.4 GB
g05/ G0.5	OpenGalaxea/G05 (g05-base)	✅	✅	11.9 GB
molmoact2/ MolmoAct2	allenai/MolmoAct2	✅	✅	21.8 GB
lingbot-vla-v2/ LingBot-VLA v2 6B	robbyant/lingbot-vla-v2-6b	✅	✅	12.8 GB

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

Tasks (LeRobot v3.0, 30 fps)

Task	Dataset	Instruction	Episodes	Frames
MugOnCoaster	aarontung/droid-mug-on-coaster	"mug on coaster"	100	41,227
CloseLaptop	aarontung/droid-close-laptop	"close laptop"	100	27,649

Common setup:

• Cameras: left exterior, right exterior and wrist (all 360x640).
• State and action: 8-D absolute joint positions [Franka arm 7, gripper 1].
• Precision: fp32 master weights with bf16 compute.

Recipes (same for every task)

Model	Cameras	Trained parts	GBS (micro-batch x accumulation, 1 GPU)	Peak LR
GR00T N1.7	left, wrist	Full: VLM backbone (LLM + vision) + action head	32 (32x1)	1e-4, cosine, 5% warmup
π0.5	left, wrist	Full (openpi), EMA 0.99	32 (32x1)	2.5e-5, cosine to 2.5e-6, 1k warmup
G0.5	left, wrist	Full (vision LR x0.1), FSDP	16 (16x1)	4e-5, cosine, 200 warmup, wd 0.03
MolmoAct2	left, right, wrist	Full	64 (32x2)	LLM 2e-5, ViT 1e-5, connector 1e-5, action expert 1e-4
LingBot-VLA v2	left, right, wrist	Full	32 (16x2)	1.25e-5, constant, Muon

The global batch sizes match the 5-epoch runs, which used 2 GPUs (per-GPU batch x 2).

Training runs

Steps and wall times are for the 5 extra epochs. The 5-epoch wall times are from the DROID-models cards.

Model	Task	Steps (+5 ep)	5-epoch run	+5 epochs (this repo)	W&B run
GR00T N1.7	MugOnCoaster	6,442	1h36m on 2x H100 PCIe	1h19m on 1x H200	ob7ln65b
GR00T N1.7	CloseLaptop	4,321	1h13m on 2x H100 PCIe	1h00m on 1x H200	c1xe5j3h
π0.5	MugOnCoaster	6,442	3h13m on 2x H100 PCIe	2h35m on 1x H200	viwljl9h
π0.5	CloseLaptop	4,321	2h17m on 2x H100 PCIe	1h47m on 1x H200	csrhwyt0
G0.5	MugOnCoaster	~12,770	4h57m on 2x H100 PCIe	3h27m on 1x H200	7s8bqwqd
G0.5	CloseLaptop	~8,570	3h12m on 2x H100 PCIe	2h23m on 1x H200	mjm8lafv
MolmoAct2	MugOnCoaster	3,221	3h55m on 2x H200 NVL	7h35m on 1x H200	rl85t3wu
MolmoAct2	CloseLaptop	2,161	2h39m on 2x H200 NVL	5h08m on 1x H200	hhlsqywh
LingBot-VLA v2	MugOnCoaster	6,442	3h23m on 2x H200 NVL	5h22m on 1x H200	wreroguc
LingBot-VLA v2	CloseLaptop	4,321	2h20m on 2x H200 NVL	3h40m on 1x H200	prd3kapz

Training logs are in the W&B project shaileshxml-nus/RoboColosseum. The 5 extra epochs are appended to the same run.

Training loss

Each value is the mean over the first or last 5% of logged steps. Columns:

• End of 5 epochs: the last 5% of the 5-epoch phase.
• Start of +5: the first 5% of the extra epochs. These values are slightly higher because the optimizer state is fresh and the LR warms up again.
• End of 10 epochs: the last 5% of the extra epochs.

Model	Task	Loss metric	End of 5 epochs	Start of +5	End of 10 epochs
GR00T N1.7	MugOnCoaster	train/loss	0.0411	0.0559	0.0281
GR00T N1.7	CloseLaptop	train/loss	0.0415	0.0565	0.0297
π0.5	MugOnCoaster	loss	0.0023	0.0023	0.0017
π0.5	CloseLaptop	loss	0.0022	0.0024	0.0016
G0.5	MugOnCoaster	fm_loss	0.0232	0.0394	0.0149
G0.5	CloseLaptop	fm_loss	0.0219	0.0356	0.0149
MolmoAct2	MugOnCoaster	train/action_flow_loss	0.0114	0.0131	0.0084
MolmoAct2	CloseLaptop	train/action_flow_loss	0.0129	0.0145	0.0108
LingBot-VLA v2	MugOnCoaster	VLA_Loss (total Loss)	—	0.0394 (0.0435)	0.0262 (0.0298)
LingBot-VLA v2	CloseLaptop	VLA_Loss (total Loss)	—	0.0409 (0.0454)	0.0265 (0.0303)

LingBot writes its losses to TensorBoard, not W&B. Its values come from the console log of the extra epochs, so there is no "end of 5 epochs" value. Training loss is not a measure of task success; the RoboColosseum evaluations measure task success.

Notes

• GR00T weight size: the GR00T folders hold fp32 weights (model-*.safetensors, 3 shards). The 5-epoch GR00T folders in DROID-models hold bf16 weights (6.5 GB). Both load the same way.
• What is not included: optimizer state. Only the final checkpoint of each run is kept.
• Code: RoboColosseum harness, scripts/train/<model>.sh, with embodiment DROID.