- NN dynamics models and control policies are effective in robotics.
- Providing sound guarantees under uncertainty is difficult, especially for closed-loop NN systems.
- Existing reachability tools are often non-differentiable, overly conservative, or too slow.
- We present a parallelizable, differentiable reachability framework in JAX.
- It handles continuous- and discrete-time systems with analytical and NN-based dynamics and controllers.
2. 先行研究と比べてどこがすごい?
- Existing reachability tools provide formal over-approximations but are often non-differentiable, overly conservative, or too slow.
- Our framework is parallelizable and differentiable, enabling integration with modern learning and online planning pipelines.
- It combines Taylor-model flowpipe construction with CROWN-style linear bound propagation.
- This unified representation preserves affine dependencies and supports GPU-batched computation and automatic differentiation.
3. 技術・手法の肝は?
- Combines Taylor-model flowpipe construction with CROWN-style linear bound propagation.
- Uses a unified representation that preserves affine dependencies.
- Supports GPU-batched computation and automatic differentiation in JAX.
- Develops a certified training method for reachability-friendly dynamics models and controllers.
- Develops a reachability-aware sampling-based MPC scheme with gradient-based refinement.
4. どうやって有効だと検証した?
- Experiments on non-prehensile manipulation and quadrotor tasks.
- Includes hardware and higher-dimensional evaluations (up to 72D).
- Demonstrates practical online planning while maintaining certified reachable-set over-approximations under bounded uncertainty.
5. 議論はある?
- The framework addresses limitations of existing reachability tools.
- It enables certified training and reachability-aware MPC.
- Experiments show practical online planning with certified over-approximations.
- Further discussions on scalability, conservativeness, or limitations are not detailed in the abstract.
6. 次に読むべき論文は?
- Taylor-model flowpipe construction
- CROWN-style linear bound propagation
- Sampling-based MPC
- Certified training for NN dynamics and controllers
- Reachability analysis for closed-loop NN systems
Neural network (NN) dynamics models and control policies achieve strong performance in robotics, but providing sound guarantees under uncertainty remains difficult, especially for closed-loop NN systems. Existing reachability tools provide formal over-approximations, yet are often non-differentiable, overly conservative, or too slow for modern learning and online planning pipelines. To address this, we present a parallelizable, differentiable reachability framework in JAX for continuous- and discrete-time systems with analytical and NN-based dynamics and controllers. Our framework combines Taylor-model flowpipe construction with CROWN-style linear bound propagation through a unified representation that preserves affine dependencies while supporting GPU-batched computation and automatic differentiation. Building on this reachability primitive, we develop (i) a certified training method that encourages reachability-friendly dynamics models and controllers, and (ii) a reachability-aware sampling-based MPC scheme with gradient-based refinement. Experiments on non-prehensile manipulation and quadrotor tasks, including hardware and higher-dimensional evaluations (up to 72D), demonstrate practical online planning while maintaining certified reachable-set over-approximations under bounded uncertainty.