NAVER LABS Europe Researchers Prove Recurrent Models Can Act as Tunable State Observers
Researchers at NAVER LABS Europe and collaborators have published a paper that gives recurrent neural networks formal guarantees for estimating the state of a dynamical system. The work, to be presented at NeurIPS in Sydney, treats RNNs through the lens of nonlinear observer theory and tests the result on a real mobile robot.
- The paper gives recurrent models formal convergence guarantees for state estimation by using nonlinear observer theory.
- Enforcing a contraction condition on the latent flow guarantees a smooth, injective map of the true state into the latent space.
- Raising a single scalar gain can drive estimation error to an arbitrarily small neighborhood of zero at an exponentially fast rate.
- On a real mobile robotic task, a provably contracting model outperformed a vanilla GRU with more stable training.
- The paper will be presented at NeurIPS in Sydney, 6 to 12 December 2026.
The problem: strong results without guarantees
Estimating the state of a dynamical system from observations is, in the authors' words, "a crucial task in robotics and control." Recurrent models have worked well in practice for this purpose, but according to the paper they often lack formal guarantees on convergence and robustness.
The paper, titled "Provable state estimation with recurrent models," is by Vincent Andrieu, Pauline Bernard, Lucas Brivadis, Laurent Praly and Christian Wolf. It is listed for the Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS), held in Sydney, Australia, from 6 to 12 December 2026.
How the theory works
The authors analyze recurrent architectures from the perspective of nonlinear observer theory. They frame RNNs as a time-discretization of continuous latent dynamical systems, known as Neural ODEs. Within this framing, they show that enforcing a contraction condition on the latent flow guarantees the existence of a smooth, injective map of the true physical state into the latent space.
The architecture then behaves as a tunable observer. According to the paper, increasing a single scalar gain that scales the latent vector field can drive the estimation error to an arbitrarily small neighborhood of zero, at an exponentially fast rate.
The work also covers decoding. The authors establish quantitative bounds on the MLP complexity needed to decode the latent state, and they extend their formal guarantees to the discrete-time residual RNN implementation.
Test on a real robot
The team validated the theory on a real mobile robotic task. There, a provably contracting model outperformed a vanilla GRU, and training was substantially more stable, according to the authors.
They also report a further observation: unconstrained GRUs converge to contracting dynamics on this task. From this, the authors suggest that contraction is an emergent property of well-trained sequence models.
What it means
For robotics and control, where state estimation underpins safe operation, the paper offers a route to recurrent models whose behavior can be backed by proofs rather than only by empirical results. The finding that standard GRUs tend toward contracting dynamics on the tested task hints that the theory may describe what well-trained models already do. The source reports results from one real robotic task, so the breadth of these findings beyond it is not established by the text.
NAVER LABS Europe: Provable state estimation with recurrent models
