Rectified Schrödinger Bridge Matching for Few-Step Visual Navigation

Abstract: Visual navigation is a core challenge in Embodied AI, requiring autonomous agents to translate high-dimensional sensory observations into continuous, long-horizon action trajectories. While generative policies based on diffusion models and Schrödinger Bridges (SB) effectively capture multimodal action distributions, they require dozens of integration steps due to high-variance stochastic transport, posing a critical barrier for real-time robotic control. We propose Rectified Schrödinger Bridge Matching (RSBM), a framework that exploits a shared velocity-field structure between standard Schrödinger Bridges ($\varepsilon=1$, maximum-entropy transport) and deterministic Optimal Transport ($\varepsilon\to 0$, as in Conditional Flow Matching), controlled by a single entropic regularization parameter $\varepsilon$. We prove two key results: (1) the conditional velocity field's functional form is invariant across the entire $\varepsilon$-spectrum (Velocity Structure Invariance), enabling a single network to serve all regularization strengths; and (2) reducing $\varepsilon$ linearly decreases the conditional velocity variance, enabling more stable coarse-step ODE integration. Anchored to a learned conditional prior that shortens transport distance, RSBM operates at an intermediate $\varepsilon$ that balances multimodal coverage and path straightness. Empirically, while standard bridges require $\geq 10$ steps to converge, RSBM achieves over 94% cosine similarity and 92% success rate in merely 3 integration steps -- without distillation or multi-stage training -- substantially narrowing the gap between high-fidelity generative policies and the low-latency demands of Embodied AI.

• The paper introduces Rectified Schrödinger Bridge Matching, which combines a learned conditional prior with entropic bridge rectification to generate eight-waypoint navigation trajectories in as few as three ODE steps without distillation.
• The method reduces velocity-target variance linearly with the regularization parameter ε while preserving the velocity-field structure, improving coarse-step integration and achieving 1.90 MSE, 0.945 cosine similarity, and 92% success at five function evaluations.
• RSBM cuts function evaluations 3.8× versus a full-budget bridge baseline, generalizes across five navigation datasets, and runs at about 50 ms per decision cycle on a Jetson Orin, although very small ε values can reduce multimodal diversity.

Overview

"Rectified Schrödinger Bridge Matching for Few-Step Visual Navigation" (2604.05673) addresses a central deployment bottleneck for generative navigation policies: diffusion- and Schrödinger Bridge–based policies capture multimodal action distributions but require many integration steps because their Brownian transport paths are high-variance and curved. The authors propose Rectified Schrödinger Bridge Matching (RSBM), a single-stage framework that introduces an entropic regularization parameter $\varepsilon \in (0,1]$ into the bridge transition kernel, interpolating between maximum-entropy Schrödinger Bridges ($\varepsilon = 1$) and deterministic optimal-transport flow matching ($\varepsilon \to 0$). The framework is anchored to a learned conditional prior that shortens the effective transport distance, enabling high-fidelity trajectory generation in as few as 3 ODE steps without distillation or multi-stage training.

Method

RSBM formulates visual navigation as conditional generative modeling: a dual-stream EfficientNet-B0/Transformer encoder maps streaming RGB observations and a goal image into a context vector $\mathbf{c} \in \mathbb{R}^{256}$; a variational prior network $g_\psi$ produces a coarse action initialization $\mathbf{a}_T$; and a conditional U-Net 1D velocity network $\mathbf{v}_\theta$ with FiLM conditioning refines $\mathbf{a}_T$ into an 8-waypoint trajectory via a probability-flow ODE.

The core construction is the $\varepsilon$-rectified conditional bridge kernel. The mean follows the standard Brownian-bridge schedule $$\boldsymbol{\mu}_t = s_t \mathbf{a}_T + (1-s_t)\mathbf{a}_0$$ with $\varepsilon = 1$0, while the variance is scaled by $\varepsilon = 1$1: $\varepsilon = 1$2. This preserves exact boundary pinning at both endpoints for any $\varepsilon = 1$3. Setting $\varepsilon = 1$4 recovers the standard Schrödinger Bridge; as $\varepsilon = 1$5 the kernel collapses to the deterministic displacement interpolant of Monge–Kantorovich optimal transport. The paper grounds this parameterization in entropic optimal transport: identifying $\varepsilon = 1$6 with the ratio of entropic regularization strengths yields a closed-form KL divergence between rectified and standard bridges of $\varepsilon = 1$7, which is 1.55 nats at the default $\varepsilon = 1$8 with $\varepsilon = 1$9.

Training uses a simulation-free conditional flow matching loss on $\varepsilon \to 0$0-prediction targets, integrated at inference with a second-order Heun solver over a Karras schedule, giving NFE $\varepsilon \to 0$1 per $\varepsilon \to 0$2 steps.

Theoretical Results

Two formal results support the design.

Velocity Structure Invariance (Theorem 1): the logarithmic derivative $\varepsilon \to 0$3 is independent of $\varepsilon \to 0$4, since the $\varepsilon \to 0$5 factors cancel exactly. Consequently, the functional form of the conditional velocity field is invariant across the entire $\varepsilon \to 0$6-spectrum, so a single velocity-network parameterization serves all regularization strengths. In practice, $\varepsilon \to 0$7 acts only as a spatial support constrictor—concentrating training samples near the deterministic interpolant—rather than changing the learning target's structure.

Velocity Variance Reduction (Proposition 1): the conditional variance of the target velocity satisfies $\varepsilon \to 0$8, so reducing $\varepsilon \to 0$9 linearly reduces stochastic variation of the training target. The paper connects this to sampling error through a standard decomposition into approximation error and discretization error, arguing that lower target variance improves regression quality and yields smoother ODE right-hand sides amenable to coarse-step solvers. Notably, this connection is stated as consistent-with rather than derived-as: the paper does not prove a tight quantitative bound linking $\mathbf{c} \in \mathbb{R}^{256}$0 to few-step accuracy, and it acknowledges that very small $\mathbf{c} \in \mathbb{R}^{256}$1 causes over-regularization and degraded multimodal diversity.

The appendix also provides a signal-to-noise analysis motivating $\mathbf{c} \in \mathbb{R}^{256}$2-prediction: unlike $\mathbf{c} \in \mathbb{R}^{256}$3-prediction (which degrades near $\mathbf{c} \in \mathbb{R}^{256}$4) or $\mathbf{c} \in \mathbb{R}^{256}$5-prediction (which degrades near $\mathbf{c} \in \mathbb{R}^{256}$6), the SNR of $\mathbf{c} \in \mathbb{R}^{256}$7-prediction is well-behaved across the full interval, and it directly minimizes ODE integration error since the solver accumulates velocity predictions.

Empirical Results

Experiments span five public datasets (HuRoN, Recon, SACSoN, SCAND, GoStanford, ~60k trajectories total under ViNT/NoMaD splits), two simulation environments (Gazebo Custom Indoor, CitySim outdoor), and real-robot trials on a quadruped with a Jetson Orin. All methods are evaluated both at default step budgets and at $\mathbf{c} \in \mathbb{R}^{256}$8 via zero-shot test-time step reduction, uniformly applied without retraining.

Method	$\mathbf{c} \in \mathbb{R}^{256}$9	NFE	MSE↓	CosSim↑	Suc.%↑
DDPM	50	50	3.80	0.820	64
FM	10	10	2.80	0.910	82
NaviBridger	10	19	1.82	0.942	88
NaviBridger	3	5	12.00	0.710	35
RSBM	3	5	1.90	0.945	92
RSBM	10	19	1.72	0.949	93

(Custom Indoor; success/collision metrics from the paper's main comparison.)

Several results stand out. Few-step superiority: RSBM at $g_\psi$0 (NFE=5) matches or exceeds NaviBridger at its full budget of $g_\psi$1 (NFE=19)—a $g_\psi$2 reduction in function evaluations—with a 92% vs. 88% success rate and +4% over NaviBridger, while achieving $g_\psi$3 lower MSE than NaviBridger at $g_\psi$4. Under zero-shot step reduction, baselines degrade sharply: NaviBridger's CosSim falls from 0.942 to 0.710 and DDPM's to 0.320, whereas RSBM saturates early (MSE 1.90 → 1.72 from $g_\psi$5 to $g_\psi$6). Cross-dataset generalization: averaged over five real-world datasets at $g_\psi$7, RSBM attains MSE 1.19 / CosSim 0.934 versus NaviBridger($g_\psi$8=10)'s 1.28 / 0.929, with the largest gains on GoStanford and SACSoN—long-range outdoor and dynamic-obstacle domains where path curvature amplifies truncation error. Prediction-target ablation: $g_\psi$9-prediction achieves 35.6% lower MSE than $\mathbf{a}_T$0-prediction and 45.7% lower than $\mathbf{a}_T$1-prediction at $\mathbf{a}_T$2, with all three converging by $\mathbf{a}_T$3, confirming the advantage is concentrated in the few-step regime. Real-robot deployment: RSBM runs at ~50 ms per decision cycle on a Jetson Orin, meeting a 4 Hz control rate, while DDPM (~350 ms) fails due to control-loop lag.

A four-way ablation disentangles the learned prior from bridge rectification: the prior alone yields MSE 5.8 (45% success); adding $\mathbf{a}_T$4-rectification lowers MSE to 1.9 ($\mathbf{a}_T$5); from Gaussian noise, RSBM still achieves MSE 4.2 versus 12.0 for standard SB ($\mathbf{a}_T$6), isolating rectification's contribution independent of prior quality. The two components are multiplicative rather than redundant. A solver ablation further shows Euler at $\mathbf{a}_T$7 already surpasses NaviBridger($\mathbf{a}_T$8=10), indicating the advantage stems from the rectified bridge geometry rather than solver choice.

Limitations and Open Questions

The paper concedes several constraints. Simulation experiments evaluate closed-loop navigation, but real-world dataset results follow the open-loop offline protocol of prior work; the preliminary real-robot validation covers only a small number of indoor scenes without a standardized benchmark or dynamic obstacles. The learned prior limits zero-shot transfer to new environments. On the theoretical side, the variance-reduction result motivates but does not formally bound few-step discretization error as a function of $\mathbf{a}_T$9; the empirical tradeoff also shows that overly small $\mathbf{v}_\theta$0 sacrifices multimodal coverage, and the choice $\mathbf{v}_\theta$1 was tuned on one environment's validation set and held fixed elsewhere—an assumption about transferability of the operating point that is not independently validated per domain. Whether the intermediate-$\mathbf{v}_\theta$2 sweet spot shifts across architectures, trajectory horizons, or out-of-distribution goals remains open.

Conclusion

RSBM unifies Schrödinger Bridges and flow matching through a single entropic parameter $\mathbf{v}_\theta$3, proving that the conditional velocity field's structure is invariant across the family while its variance scales linearly with $\mathbf{v}_\theta$4. Combined with a learned conditional prior, this yields a single-stage generative policy that reaches 94.5% cosine similarity and 92% success rate in 3 ODE steps, matching full-budget bridge baselines with $\mathbf{v}_\theta$5 fewer function evaluations and no distillation. The framework provides a direct latency–quality knob for embodied deployment, though its real-world evidence remains limited to open-loop benchmarks and small-scale indoor robot trials.