Research map / Architectures and efficiency / RNNs and normalization
Model-Based Control with Sparse Neural Dynamics
Ziang Liu, Genggeng Zhou, Jeff He, Tobia Marcucci, Li Fei-Fei, Jiajun Wu, Yunzhu Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MIP-based planning instead of MPPI with the learned dynamics model | Sparse Dynamics · Object pushing, object sorting, rope manipulation | open-loop planning performance | Improved |
| Use sparsified dynamics models with MIP-based control rather than the evaluated RL baselines | Closed-loop control · CartPole-v1, Reacher-v4 | closed-loop control performance | Improved |
| Reduce the number of ReLUs while retaining a reasonably accurate dynamics model | Sparse Dynamics · Object pushing, object sorting, rope manipulation, CartPole-v1, Reacher-v4 | closed-loop control performance | Improved |
| Reduce the number of ReLUs in the dynamics model | Sparse Dynamics · Object pushing, object sorting, rope manipulation | long-horizon prediction performance | No clear effect |
| Replace nonlinear units with identity mappings as well as removing units | Dynamics models at matched ReLU counts · Object pushing, object sorting, rope manipulation | prediction error | Improved |
| Use the proposed nonlinearity sparsification instead of channel pruning | Object pushing · Object pushing | closed-loop control performance | Improved |
| Sparsify the model to two ReLUs | Overparameterized MLP for piecewise affine recovery · Piecewise affine functions | ground-truth system recovery | Improved |
| Use channel pruning instead of the proposed sparsification | Overparameterized MLP for piecewise affine recovery · Piecewise affine functions | ground-truth system recovery | Hurt |
| Use a simplified learned dynamics model as the forward model for PPO | PPO using learned dynamics · CartPole-v1 | closed-loop control performance | Hurt |
Closest papers by meaning
- DDPNOpt: Differential Dynamic Programming Neural Optimizer (ICLR 2021 · similarity 0.88)
- Deep learning for continuous-time stochastic control with jumps (NeurIPS 2025 · similarity 0.84)
- BaB-ND: Long-Horizon Motion Planning with Branch-and-Bound and Neural Dynamics (ICLR 2025 · similarity 0.84)
- Physics-Informed Neural Network Policy Iteration: Algorithms, Convergence, and Verification (ICML 2024 · similarity 0.84)
- MAST: model-agnostic sparsified training (ICLR 2025 · similarity 0.84)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.84)
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.
