To improve zero-shot goal generalization of behavior cloning (BC), the authors propose TD-Aug.
Imagined goals from geometric horizon models boost behavior cloning generalization
TD-Aug samples synthetic future states from a temporal difference trained world model, raising zero-shot success on OGBench stitching tasks from 31% to 79%.
Big Tech
Daniel Lawson · Dhruv Sreenivas · Adriana Hugessen · Khimya Khetarpal · Glen Berseth
Mila · Université de Montréal · Google DeepMind
Research Digest··2 min read
Lawson and colleagues propose TD-Aug, a data augmentation method that trains a geometric horizon model with temporal difference flows and uses its samples as imagined goals for goal-conditioned behavior cloning.
Why this paper
From Google DeepMind and 2 others
In one line
Goal-conditioned behavior cloning generalizes to unseen compositional tasks when trained on goals sampled from a temporal-difference geometric horizon model.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§