Views
No views yet
CoffeePressButton demonstrations
(54 episodes, cosmos_predict2_2b_480p_robocasa_50_demos_per_task experiment config),
trained to 100,000 iterations (batch_size effective 4).RoboCasa-Cosmos-Policy/success_only)run_robocasa_eval.py, matched train layout/style, 5 trials): see repo eval logs.pt (from DCP via dcp_to_torch_save), load with
cosmos_policy.experiments.robot.robocasa.run_robocasa_eval (--ckpt_path) or
cosmos_policy inference config cosmos_predict2_2b_480p_robocasa_50_demos_per_task__inference.dataset_statistics.json (proprio/action normalization stats) and
t5_embeddings.pkl (task text T5 embeddings) required for inference.