Worked example: packaging a multimodal robot dataset
範例
A reseller has a multimodal robot dataset with these givens: 30 Hz RGB camera, 10 Hz LiDAR, 200 Hz IMU, and 100 Hz joint states, captured across 500 episodes of a pick-and-place task. Each episode lasts about 20 seconds. The buyer's stated use is training an imitation-learning policy that maps observations to actions. The reseller must choose a storage and format approach. Assume the buyer will load whole episodes, not run cross-episode statistics.
示範講解
Step 1 — Identify the profile. The dataset is multimodal, high-rate, and episode-structured, and the buyer's use is sequence learning. That profile points toward an episode-based layout rather than a columnar one, because the buyer needs each run intact. Step 2 — Check synchronization feasibility. The streams run at different rates (30, 10, 200, 100 Hz), so a common timeline is required before packaging. The pipeline resamples or indexes each stream against a shared clock so that every action label can be tied to the sensory state at that moment. Step 3 — Choose the format. An episode-based layout stores each run as a self-contained unit with its streams and labels, which matches the buyer's access pattern. A manifest records the sensor list, rates, units, and episode count so the buyer can verify the package without opening every file. Step 4 — Check the delivery consequence. Because the layout matches the buyer's use, the buyer can load an episode and train directly. Had the reseller chosen a columnar layout instead, the buyer would have to reconstruct sequences from scattered rows before training could begin.