Excited to introduce DEPS (Discovery of GenEralizable Parameterized Skills) at #NeurIPS2025! DEPS learns interpretable parameterized skills that drastically improve generalisation to unseen tasks, especially in data-constrained settings and on out-of-distribution tasks. (1/n)

@haotiannnnnnnnn @calvinyluo @yidingjiang For more, please check out our: Website: https://t.co/TUNfZyhxUH Arxiv: https://t.co/hRnlvIlw8R Code: https://t.co/ioQj4psLlk See y’all at NeurIPS! Feel free to message me here or at vedantgupta@gmail.com with questions:) (n/n)

Bonus: DEPS discovers interpretable skills! Visualisations are on our website ⬇️ I had a great time working on DEPS with my amazing collaborators @haotiannnnnnnnn, @calvinyluo, @yidingjiang and George Konidaris! (7/n)

The result is superior generalization across diverse evaluation regimes: E.g. 2.3x higher average success than baselines on out-of-distribution LIBERO tasks. 2.4x better with 3-shot learning 4x better with limited pretraining In short, DEPS learns skills that transfer. (6/n)

To make this work, we make several key architectural choices. E.g., we view parameterized skills as low-dimensional trajectory manifolds. Trajectories can be indexed into with a scalar → DEPS compresses the robot’s state to 1D before feeding it to the low-level policy (5/n)

To address this, prior work uses "staged" training, e.g. using VLMs to pre-segment trajectories. If the segmentation goes wrong, you’re left with bad skills. OTOH, DEPS learns parameterized skills in an end-to-end fashion - single training process, no pretrained models. (4/n)

The challenge? Naively training hierarchical policies (discrete skill → continuous params → actions) doesn’t really work… There are just too many ways to fit the training data without learning the nice, reusable skills we’re looking for. (3/n)

What's a parameterized skill👀 Key idea: parameterized skills = discrete behaviors + continuous parameters Think: pick(x,y,z) where the skill is "pick" but (x,y,z) specify where Here’s a pick skill discovered by DEPS. Different continuous parameters pick different objects (2/n)