Linslab's picture
Add level2_dual_arm multi-task PI0.5 LoRA (14 level-2 tasks, 50k steps, mean success 0.120)
e8d4dfc verified
Raw History Blame Contribute Delete
2.18 kB
{
"suite": "level2_dual_arm",
"suite_id": "l2dual",
"description": "all level-2 dual-arm tasks (gripper + dexterous hand)",
"level": "level_2",
"step": 50000,
"campaign": "pi05_lv2sub_20261029",
"tasks": 14,
"mean_success_rate": 0.11999999999999998,
"episodes_per_task": 50,
"episode_seeds": "1000-1049",
"per_task": [
{
"task": "DualArmLiftBall-v1",
"success_rate": 0.28,
"successes": 14,
"return_mean": 31.4384
},
{
"task": "DualArmBookPickPlace-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 536.0200600000001
},
{
"task": "DualArmPegInHole-v1",
"success_rate": 0.08,
"successes": 4,
"return_mean": 45.02074
},
{
"task": "DualArmCleanupPotEnv-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 144.15777999999997
},
{
"task": "DualArmClothAlign-v1",
"success_rate": 0.26,
"successes": 13,
"return_mean": 24.618180000000002
},
{
"task": "DualArmPushObject-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 0.04298
},
{
"task": "DualArmStraightenRope-v1",
"success_rate": 0.2,
"successes": 10,
"return_mean": 51.66449999999999
},
{
"task": "DualArmOpenLaptop-v1",
"success_rate": 0.14,
"successes": 7,
"return_mean": 1533.2114400000005
},
{
"task": "DualArmThrowBlock-v1",
"success_rate": 0.32,
"successes": 16,
"return_mean": 2062.5982
},
{
"task": "DualArmJugglingBall-v1",
"success_rate": 0.12,
"successes": 6,
"return_mean": -4807.480519999998
},
{
"task": "DualArmOpenBottle-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 2186.440520000001
},
{
"task": "DualArmPlaceRopeOverRod-v1",
"success_rate": 0.16,
"successes": 8,
"return_mean": 268.38523999999995
},
{
"task": "DualArmStuffCloth-v1",
"success_rate": 0.12,
"successes": 6,
"return_mean": 30.693659999999994
},
{
"task": "DualArmPouringTwoCups-v1",
"success_rate": 0.0,
"successes": 0,
"return_mean": 2262.00242
}
],
"note": "Only the final checkpoint (step 50000) was evaluated; 50 episodes per task, the same seeds for every task and suite."
}