Research map / Architectures and efficiency / RNNs and normalization
Learning Options for Compositional Motor Control with Adapter Banks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CABRA, a shared LSTM core modulated by a bank of ten residual adapters selected by a CompILE-style discrete latent code learned from expert demonstrations of closed-loop biomechanical arm control, and shows emergent low-rank adapter perturbations and out-of-distribution compositional transfer.
Problem
How motor primitives implemented as low-rank perturbations of a shared recurrent network can be learned end-to-end from demonstrations, and whether the resulting primitives recombine to support out-of-distribution compositional motor transfer.
Approach
A CompILE-style encoder segments compound expert demonstrations into discrete codes that select residual adapters (full-rank, unconstrained) modulating a shared LSTM core; low-rank structure emerges from training, and a frozen-network soft policy over codes sequences the adapters for novel trajectories.
Key results
- Student reaches hand L1 approximately 0.005 on the two-petal transfer versus about 0.04 for the multitask teacher, an order-of-magnitude gap.
- Learned adapter perturbations have effective rank about 5-6 each, while the core hidden-state manifold has effective rank about 30 out of 256 units.
- Student task subspaces show mean principal-subspace overlap 0.18 versus 0.69 for the teacher.
Surprising findings
- Despite full-rank adapter parameterization, individual adapter perturbations converge to low effective rank (about 5-6).
- Teacher and student dynamics differ: the teacher's task representations overlap heavily in one shared subspace, while the student's are highly separated.
Limitations
- Training uses task-identity supervision on the encoder's segment-1 representation, so the purely unsupervised variant is untested.
- Student is evaluated only on a single biomechanical arm environment with five compound tasks and two OOD shapes.
- Forced-code analysis shows partial sharing among retraction codes, so retraction primitives are less one-to-one than extension primitives.
Experimental setup
Datasets: closed-loop teacher-generated demonstrations of compound biomechanical movements
Benchmarks: held-out compound task hand-trajectory L1 two-petal OOD trajectory butterfly OOD trajectory
Metrics: hand trajectory L1 error effective rank (95% variance) principal-subspace overlap DSA distance
Baselines: task-input-conditioned multitask RNN teacher single shared adapter ablation no-adapter ablation
Builds on: CompILE DIAYN thalamocortical low-rank perturbation framework LoRA Dynamical Similarity Analysis
Closest papers by meaning
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.87)
- Learning Dynamics of RNNs in Closed-Loop Environments (NeurIPS 2025 · similarity 0.86)
- World Models as Reference Trajectories for Rapid Motor Adaptation (NeurIPS 2025 · similarity 0.86)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.86)
- Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.86)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.