Add config.json for download tracking
#2
by
moojink - opened
- config.json +82 -0
config.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "cosmos-policy-planning",
|
| 3 |
+
"architecture": "diffusion-transformer",
|
| 4 |
+
"base_model": "nvidia/Cosmos-Policy-ALOHA-Predict2-2B",
|
| 5 |
+
"num_parameters": "2B",
|
| 6 |
+
|
| 7 |
+
"input_spec": {
|
| 8 |
+
"text": {
|
| 9 |
+
"type": "string",
|
| 10 |
+
"description": "Natural language task description"
|
| 11 |
+
},
|
| 12 |
+
"images": {
|
| 13 |
+
"format": "RGB",
|
| 14 |
+
"resolution": [224, 224],
|
| 15 |
+
"views": ["top_down", "left_wrist", "right_wrist"]
|
| 16 |
+
},
|
| 17 |
+
"proprioception": {
|
| 18 |
+
"dim": 14,
|
| 19 |
+
"components": ["left_arm_joints", "right_arm_joints"],
|
| 20 |
+
"joints_per_arm": 7
|
| 21 |
+
},
|
| 22 |
+
"actions": {
|
| 23 |
+
"dim": 14,
|
| 24 |
+
"horizon": 50,
|
| 25 |
+
"description": "Candidate action sequence to evaluate"
|
| 26 |
+
}
|
| 27 |
+
},
|
| 28 |
+
|
| 29 |
+
"output_spec": {
|
| 30 |
+
"future_proprioception": {
|
| 31 |
+
"dim": 14
|
| 32 |
+
},
|
| 33 |
+
"future_images": {
|
| 34 |
+
"resolution": [224, 224],
|
| 35 |
+
"views": 3
|
| 36 |
+
},
|
| 37 |
+
"value": {
|
| 38 |
+
"dim": 1,
|
| 39 |
+
"description": "Expected cumulative reward for action sequence"
|
| 40 |
+
}
|
| 41 |
+
},
|
| 42 |
+
|
| 43 |
+
"diffusion_config": {
|
| 44 |
+
"denoising_steps": 10,
|
| 45 |
+
"sigma_min": 4.0,
|
| 46 |
+
"sigma_max": 80.0
|
| 47 |
+
},
|
| 48 |
+
|
| 49 |
+
"planning_config": {
|
| 50 |
+
"ensemble_world_model_queries": 3,
|
| 51 |
+
"ensemble_value_queries": 5,
|
| 52 |
+
"total_predictions_per_action": 15,
|
| 53 |
+
"best_of_n_search": 8
|
| 54 |
+
},
|
| 55 |
+
|
| 56 |
+
"training": {
|
| 57 |
+
"dataset": "ALOHA policy rollouts",
|
| 58 |
+
"num_episodes": 648,
|
| 59 |
+
"hardware": "8x H100",
|
| 60 |
+
"batch_split": {
|
| 61 |
+
"policy": 0.1,
|
| 62 |
+
"world_model": 0.45,
|
| 63 |
+
"value_function": 0.45
|
| 64 |
+
}
|
| 65 |
+
},
|
| 66 |
+
|
| 67 |
+
"benchmark_results": {
|
| 68 |
+
"put_candies_in_bowl": 0.60,
|
| 69 |
+
"put_candy_in_ziploc_bag": 0.84,
|
| 70 |
+
"average": 0.72,
|
| 71 |
+
"improvement_over_base": 0.125
|
| 72 |
+
},
|
| 73 |
+
|
| 74 |
+
"inference": {
|
| 75 |
+
"precision": "bf16",
|
| 76 |
+
"latency_seconds": 4.9,
|
| 77 |
+
"recommended_gpus": 8
|
| 78 |
+
},
|
| 79 |
+
|
| 80 |
+
"robot_platform": "ALOHA 2 (ViperX 300 S dual arms)",
|
| 81 |
+
"control_frequency_hz": 25
|
| 82 |
+
}
|