Linslab's picture
Add level2_dexterous multi-task PI0.5 LoRA (14 level-2 tasks, 50k steps, mean success 0.210)
29dbe7c verified
Raw History Blame Contribute Delete
2.15 kB
{
"suite": "level2_dexterous",
"suite_id": "l2dex",
"description": "all level-2 dexterous-hand tasks (single-arm + dual-arm)",
"level": "level_2",
"step": 50000,
"campaign": "pi05_lv2sub_20261029",
"tasks": 14,
"mean_success_rate": 0.21000000000000002,
"episodes_per_task": 50,
"episode_seeds": "1000-1049",
"per_task": [
{
"task": "GraspTwoBlocks-v1",
"success_rate": 0.64,
"successes": 32,
"return_mean": 213.11556000000002
},
{
"task": "ReorientPen-v1",
"success_rate": 0.26,
"successes": 13,
"return_mean": 107.33642
},
{
"task": "DexOpenDoor-v1",
"success_rate": 0.26,
"successes": 13,
"return_mean": 336.71574000000015
},
{
"task": "HammerNailOut-v1",
"success_rate": 0.22,
"successes": 11,
"return_mean": 90.09622000000002
},
{
"task": "ClothWipe-v1",
"success_rate": 0.1,
"successes": 5,
"return_mean": 385.02892000000014
},
{
"task": "RollingPin-v1",
"success_rate": 0.2,
"successes": 10,
"return_mean": 271.1742999999999
},
{
"task": "RopeBridgeDex-v1",
"success_rate": 0.16,
"successes": 8,
"return_mean": 308.84354
},
{
"task": "DualArmOpenLaptop-v1",
"success_rate": 0.08,
"successes": 4,
"return_mean": 1571.8813399999997
},
{
"task": "DualArmThrowBlock-v1",
"success_rate": 0.2,
"successes": 10,
"return_mean": 2143.21196
},
{
"task": "DualArmJugglingBall-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": -5159.3756
},
{
"task": "DualArmOpenBottle-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 2093.2040599999996
},
{
"task": "DualArmPlaceRopeOverRod-v1",
"success_rate": 0.58,
"successes": 29,
"return_mean": 368.10752
},
{
"task": "DualArmStuffCloth-v1",
"success_rate": 0.24,
"successes": 12,
"return_mean": 28.31748000000001
},
{
"task": "DualArmPouringTwoCups-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 2654.445779999999
}
],
"note": "Only the final checkpoint (step 50000) was evaluated; 50 episodes per task, the same seeds for every task and suite."
}