NeurIPS paper learns compositional motor options with adapter banks, cutting generalization error up to 10x
sreejan_kumar · x · 2026-09-25
A paper by Sreejan Kumar, Marcelo Mattar and Lea Duncker accepted to NeurIPS 2026 translates the neuroscience idea that motor primitives are low-rank perturbations of a shared recurrent network into an end-to-end architecture: a shared recurrent core modulated by a bank of residual adapters, each selected via a discrete latent code.
- Trained on closed-loop biomechanical control, adapters emergently develop low-rank perturbations with no architectural rank constraint, placing task representations in disparate subspaces.
- A simple high-level policy over the learned options, optimized while the network is frozen, sequences adapters to produce novel out-of-distribution movements.
- Generalization error improves on a task-input-conditioned multitask baseline by up to an order of magnitude.
The author calls it a precursor to what he hopes will be one of the biggest papers of his career.
More from Embodied
- Disney Research shows how it trains robots to dance at monthly robotics AI paper club — DominiqueCAPaul · 2026-09-25
- NVIDIA pitches GPU-accelerated physics simulation as the training ground for real-world robots — LexiLove · 2026-09-25
- Community vote crowns Physical Intelligence as the robotics lab with top talent density — ZeYanjie · 2026-09-25
- AMD and Perplexity team up to bring Perplexity Portable Computer to Ryzen AI Halo — AnushElangovan · 2026-09-25
- Google Glass founder: glasses that block eye contact will never go mainstream — Scobleizer · 2026-09-25
- 19 Unitree Humanoids Dance With 120 Performers Before 10,000+ in Shanghai's Record Live Show — rohanpaul_ai · 2026-09-25