File size: 4,459 Bytes
9ecc04a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 | [
{
"name": "Transition Dataset",
"type": "Dataset Pattern",
"focus": "One-step dynamics",
"best_for": [
"Dynamics learning",
"Control",
"Simple environments"
],
"required": [
"observation_t",
"action_t",
"observation_t+1"
],
"optional": [
"reward",
"done",
"goal",
"timestamp"
],
"risk": "Frame-level random splits can leak nearly identical states across train and test."
},
{
"name": "Episode Dataset",
"type": "Dataset Pattern",
"focus": "Sequential dynamics",
"best_for": [
"Rollouts",
"Memory",
"Long-horizon modeling"
],
"required": [
"episode_id",
"observations[]",
"actions[]"
],
"optional": [
"rewards[]",
"terminal",
"instruction",
"metadata"
],
"risk": "Broken episode boundaries can destroy temporal structure."
},
{
"name": "Multimodal Episode",
"type": "Dataset Pattern",
"focus": "Cross-modal world state",
"best_for": [
"Robotics",
"Embodied AI",
"Physical AI"
],
"required": [
"video",
"actions",
"robot_state"
],
"optional": [
"depth",
"audio",
"language",
"force",
"pose"
],
"risk": "Unsynchronized modalities create false transition errors."
},
{
"name": "One-Step Benchmark",
"type": "Benchmark Pattern",
"focus": "Immediate next-state prediction",
"best_for": [
"Fast iteration",
"Architecture debugging"
],
"required": [
"held-out transitions",
"prediction metric"
],
"optional": [
"uncertainty",
"per-task breakdown"
],
"risk": "Can hide catastrophic long-horizon drift."
},
{
"name": "Multi-Horizon Benchmark",
"type": "Benchmark Pattern",
"focus": "Error growth over time",
"best_for": [
"Rollouts",
"Planning",
"Simulation"
],
"required": [
"horizons",
"rollout protocol",
"per-horizon metric"
],
"optional": [
"drift ratio",
"consistency score"
],
"risk": "Averaging across horizons can conceal where failure begins."
},
{
"name": "Action Fidelity Benchmark",
"type": "Benchmark Pattern",
"focus": "Correct action consequences",
"best_for": [
"Robotics",
"Interactive worlds",
"Agents"
],
"required": [
"paired actions",
"ground-truth transitions"
],
"optional": [
"counterfactual actions",
"action sensitivity"
],
"risk": "Visual similarity alone does not prove correct causal response."
},
{
"name": "OOD Generalization Split",
"type": "Split Strategy",
"focus": "Distribution shift",
"best_for": [
"Deployment",
"Robotics",
"Robustness"
],
"required": [
"held-out environments or tasks"
],
"optional": [
"held-out objects",
"held-out embodiments"
],
"risk": "Random splits can dramatically overestimate generalization."
},
{
"name": "Planning Utility Benchmark",
"type": "Benchmark Pattern",
"focus": "Downstream decision quality",
"best_for": [
"Agents",
"MPC",
"Model-based RL"
],
"required": [
"planner",
"task success metric",
"world-model rollouts"
],
"optional": [
"regret",
"return",
"sample efficiency"
],
"risk": "Prediction scores may not correlate with better decisions."
},
{
"name": "Control Benchmark",
"type": "Benchmark Pattern",
"focus": "Closed-loop task performance",
"best_for": [
"Robotics",
"Autonomous systems"
],
"required": [
"environment",
"policy/controller",
"success metric"
],
"optional": [
"safety violations",
"energy",
"path efficiency"
],
"risk": "Offline prediction quality may not transfer to closed-loop control."
},
{
"name": "Efficiency Benchmark",
"type": "Benchmark Pattern",
"focus": "Operational cost",
"best_for": [
"Real-time systems",
"Large search spaces"
],
"required": [
"latency",
"memory",
"throughput"
],
"optional": [
"energy",
"rollouts/sec",
"cost"
],
"risk": "A strong model may be unusable if rollouts are too slow for planning."
}
] |