QZWang/robocasa-human300-step559660

RoboCasa Human300 Action-as-Patch — step_559660

Human300-trained Action-as-Patch checkpoint used as the initialization for later continuation experiments.

Checkpoint

• File: step_559660.pt
• Size: 7,751,442,617 bytes
• SHA256: 7877dd2546d7e2fe489aab79f84476266724b7392d48fe2f2861e9752cd4dccb
• Training data: RoboCasa365 Human300 pretraining set
• Training: 5 epochs, per-GPU batch 16, gradient accumulation 1, bf16, learning rate 1e-4, weight decay 1e-2

RoboCasa evaluation

Protocol: RoboCasa 1.0.1, split=pretrain, replan=16, 10 denoise steps, 50 trials per task, 2,500 episodes total.

Suite	Success	Episodes	SR	Aborted
Atomic Seen	708	900	78.667%	0
Composite Seen	371	800	46.375%	0
Composite Unseen	110	800	13.750%	0
Task-weighted overall	1,189	2,500	47.560%	0

Code

https://github.com/TianhengW/robocasa-norm-step060000

The repository is currently private. The checkpoint format requires the matching Action-as-Patch model and RoboCasa preprocessing configuration.