Research map / Probabilistic and scientific modeling / Dynamical systems

Learning Dynamics under Environmental Constraints via Measurement-Induced Bundle Structures

ICML 2025 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent-space control and planning · 0 citations (OpenAlex)

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 a fiber-bundle geometric framework with measurement-aware control barrier functions (mCBFs) and Neural ODEs for learning unknown dynamics under local, uncertain environmental constraints, evaluated on simulated worm, Franka arm, and quadrotor tasks.

Problem

Learning unknown dynamics under environmental constraints when constraint information is only locally available through uncertain sensor measurements, where existing methods need global constraints or probabilistic filtering.

Approach

Treats the state-measurement relation as a fiber bundle over the state space, defines measurement-adapted CBFs whose safety margin shrinks with measurement error, and learns continuous-time dynamics with Neural ODEs under these certificates.

Key results

• Ours achieves 96.3% success rate vs. 84.0% for Neural-CBF and 89.7% for DataFilter in the main table.
• Ours reports 99.3% constraint satisfaction, versus 99.7% for GPMPC, which is the highest.
• Removing mCBFs drops success rate from 96.3% to 45.7% in the ablation.

Measured effects reported in the paper

Surprising findings

• The Lie group symmetry ablation shows only a mild drop (96.3% to 85.7%), and the paper calls it complementary rather than critical.
• The paper's own Neural ODE ablation (w/o mCBF) yields the largest degradation, yet the theoretical guarantees are stated only as sketches.

Limitations

• Current implementation struggles with highly stochastic dynamics and complex environmental uncertainty (stated by authors).
• Theoretical proofs rely on loosely justified steps such as Lipschitz assumptions and probabilistic bounds that are only sketched.
• Evaluation is entirely in simulation with self-defined metrics and sensing noise models.

Experimental setup

Datasets: simulated trajectories from Genesis physics engine (worm MPM, Franka arm, quadrotor)

Benchmarks: soft worm navigation simulation Franka arm manipulation simulation quadrotor navigation simulation

Metrics: success rate path length constraint satisfaction rate minimum safety margin control magnitude control smoothness

Baselines: Neural-CBF SafeLearn SafetyNet SafeTrack StructCBF PNDS GEM GeoPath GPMPC ALMPC SafeRL RobustSafe

Builds on: control barrier functions Neural ODE soft actor-critic

Closest papers by meaning

• Learning Physics Informed Neural ODEs with Partial Measurements (AAAI 2025 · similarity 0.85)
• Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds (NeurIPS 2026 · similarity 0.85)
• Safely Learning Controlled Stochastic Dynamics (NeurIPS 2025 · similarity 0.84)
• Geometric Contact Flows: Contactomorphisms for Dynamics and Control (ICML 2025 · similarity 0.84)
• A Differential and Pointwise Control Approach to Reinforcement Learning (NeurIPS 2025 · similarity 0.83)
• Memory-Augmented Potential Field Theory: A Framework for Adaptive Control in Non-Convex Domains (NeurIPS 2025 · similarity 0.83)