From five demonstrations
to a curated training set.
Explore how a target demonstration set retrieves related robot experience, then inspect the measured policy result.
The query: five target demonstrations
These five target demonstrations are used for both retrieval and training. The target task is demonstrated, not unseen or zero-shot. Playing a video changes playback only; the retrieval query remains the complete five-demo set.
Query selection receipt
Choose a retrieval method
Rankings are precomputed on Spark using cached representations. DTW aligns sequences; it is not an approximate nearest-neighbor vector lookup. No live model calls occur here.
Inspect the source candidates
Ranking provenance and ordering
Compare the leading candidates
Independent top-three lists for the same five-demo query and experiment. Selecting either method below changes its comparison column.
What happened after training?
Compare target-only training with STRAP data augmentation. These results are independent of the retrieval method selected above.
Experiment details
- Target demonstrations
- Five demonstrations of the selected task from LIBERO-10.
- Retrieval database
- 4,500 demonstrations across 90 tasks in LIBERO-90, with related objects and skills.
- Model and training
- A behavior-cloning Transformer trained on the target demonstrations, either alone or together with data selected by STRAP using visual subsequence DTW.
- Evaluation
- 50 simulator trials per reset setting, using one training seed for each experiment.