Organizations: School of Artificial Intelligence, Jilin University, Changchun 130012, China · College of Computer Science, Chongqing University, Chongqing 400044, China · Department of Computer Science, University of Liverpool, Liverpool L69 3BX, UK · Changchun GenY Technology Co., Ltd., Changchun, China
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 (ε=1, maximum-entropy transport) and deterministic Optimal Transport (ε→0, as in Conditional Flow Matching), controlled by a single entropic regularization parameter ε. We prove two key results: (1) the conditional velocity field's functional form is invariant across the entire ε-spectrum (Velocity Structure Invariance), enabling a single network to serve all regularization strengths; and (2) reducing ε 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 ε that balances multimodal coverage and path straightness. Empirically, while standard bridges require ≥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.
Figures & tables
Fig. 1: Denoising progression on two toy trajectories (star patrol and figure-8 loop). At t=T all methods start as unstructured point clouds (Gaussian noise for DDPM/FM; learned prior for NaviBridger/RSBM). By k=2 (NFE = 3), only RSBM closely matches the GT, while baselines require k≥10 to converge. At k=30 (NFE = 59), all methods have converged, confirming the quantitative trends in Table III .
Fig. 2: Overview of the RSBM framework. Left : A dual-stream EfficientNet-B0 vision encoder fϕ (§ III-A ) extracts observation and goal features, which are fused via positional encoding and self-attention into a context vector c∈R256 . Center : A learned variational prior network gψ (§ III-A ) produces a coarse action prior aT . Right : A conditional U-Net 1D velocity network vθ (§ III-C ) with FiLM conditioning iteratively refines aT into the output trajectory a0 via an ε -Rectified Schrödinger Bridge (§ III-B – III-D ). The three trajectory snapshots illustrate the progressive refinement: from scattered prior waypoints (green), through partially converged intermediate states (orange), to the final high-fidelity trajectory (red) closely matching the ground truth (dashed gray).
Custom Indoor
CitySim (Outdoor)
Method
k
NFE
MSE ↓
CosSim ↑
FDE ↓
Col.% ↓
Suc.% ↑
MSE ↓
CosSim ↑
FDE ↓
Col.% ↓
Suc.% ↑
ViNT [ 9 ] (1-shot)
1
1
6.50
0.720
2.85
1.58
28
8.20
0.650
4.50
0.41
38
NoMaD [ 19 ]
20
20
3.60
0.820
1.95
1.32
32
5.80
0.740
3.20
0.34
52
NoMaD [ 19 ]
3
6
8.40
0.610
3.75
2.25
18
10.50
0.540
5.20
0.55
22
DDPM [ 20 ]
50
50
3.80
0.820
2.05
0.98
64
5.50
0.750
3.10
0.35
50
DDPM [ 20 ]
3
6
14.80
0.320
6.10
3.60
6
16.20
0.280
7.80
0.72
4
TABLE I: Comprehensive comparison. Each method is shown at its default k and at k=3 , revealing degradation under zero-shot step reduction. Gray rows : k=3 variants. RSBM is shown at both k=3 and k=10 to demonstrate early saturation. Bold : best overall per column.
Fig. 3: Effect of ε . ε=0.5 (red) provides stable low- k performance.
Fig. 4: Quality–cost Pareto frontier. Each marker represents a method at a given sampling budget ( k ). (a) CosSim vs. NFE; (b) Success Rate vs. NFE. RSBM at k=3 (NFE =5 ) lies on the favorable frontier region, providing strong quality at substantially lower evaluations.
Action MSE ↓
CosSim ↑
Method
k
HuRoN
Recon
SACSoN
SCAND
GoStan.
Avg
HuRoN
Recon
SACSoN
SCAND
GoStan.
Avg
DDPM [ 20 ]
10
1.38
1.19
1.74
0.72
4.12
1.83
0.880
0.900
0.860
0.920
0.780
0.868
DDPM [ 20 ]
3
8.50
6.20
9.80
4.60
15.30
8.88
0.350
0.420
0.280
0.480
0.210
0.348
FM [ 39 ]
10
1.10
0.97
1.55
0.65
3.65
1.58
0.910
0.920
0.885
0.935
0.820
0.894
FM [ 39 ]
3
3.20
2.60
4.10
1.80
7.50
3.84
0.720
0.740
0.670
0.790
0.560
0.696
NaviBridger [ 31 ]
10
0.27
0.88
1.43
0.59
3.22
1.28
0.955
0.935
0.920
0.960
0.875
0.929
TABLE II: Per-dataset generalization. Action MSE ↓ and CosSim ↑ across five diverse real-world datasets. RSBM( k=3 ) consistently matches or exceeds NaviBridger( k=10 ); NaviBridger( k=3 ) degrades significantly across all domains. Bold : best per column among k=3 methods.
k=3 (NFE=5)
k=10 (NFE=19)
Target
MSE ↓
CosSim ↑
FDE ↓
MSE ↓
CosSim ↑
FDE ↓
ϵ -pred
3.50
0.895
1.42
1.95
0.940
0.88
x0 -pred
2.95
0.920
1.15
1.80
0.946
0.84
v -pred (Ours)
1.90
0.945
0.80
1.72
0.949
0.78
TABLE III: Prediction target ablation (RSBM, ε=0.5 , Custom Indoor). Bold : best per column.
Fig. 5: Qualitative trajectory comparison across eight challenging scenarios ( 2×4 grid, k=3 , NFE =5 ). Top row : four indoor/structured environments. Bottom row : four large-scale environments. Baselines collide early (×); faint dotted lines show invalid ghost continuations. RSBM (green) remains collision-free and closely tracks the ground truth (dashed gray).
Method
k
MSE ↓
CosSim ↑
Suc.% ↑
ViNT (1-shot)
1
6.50±0.15
0.720±0.005
28±2
NoMaD
20
3.60±0.12
0.820±0.004
32±3
DDPM
50
3.80±0.14
0.820±0.005
64±3
FM
10
2.80±0.10
0.910±0.003
82±2
NaviBridger ( k =10)
10
1.82±0.08
0.942±0.003
88±2
NaviBridger ( k =3)
3
12.00±0.45
0.710±0.015
35±4
TABLE IV: Mean ± std over 3 random seeds (Custom Indoor). Bold : best per column.
Bridging high-level semantic understanding with low-level physical control remains a persistent challenge in embodied intelligence, stemming from the fundamental spatiotemporal scale mismatch between cognition and action. Existing generative VLA policies typically adopt a "Generation-from-Noise" paradigm, which disregards this disparity, leading to representation inefficiency and weak condition alignment during optimization. In this work, we propose ResVLA, an architecture that shifts the paradigm to "Refinement-from-Intent." Recognizing that robotic motion naturally decomposes into global intent and local dynamics, ResVLA utilizes spectral analysis to decouple control into a deterministic low-frequency anchor and a stochastic high-frequency residual. By anchoring the generative process on the predicted intent, our model focuses strictly on refining local dynamics via a residual diffusion bridge. Extensive simulation experiments show that ResVLA achieves competitive performance, strong robustness to language and robot embodiment perturbations, and faster convergence than standard generative baselines. ResVLA also demonstrates strong performance in real-world robot experiments.
Yiming Zhong, Yaoyu He, Zemin Yang +5
ShanghaiTech University, Shanghai, China · Morphic Robotics, Shenzhen, China · The Chinese University of Hong Kong, Hong Kong, China.
By relying on independent couplings from uninformative Gaussian priors, standard diffusion and flow matching models are forced to learn complex, high-cost vector fields to reach the physical action space. Generative models excel at capturing multimodal behaviors for robotic Learning from Demonstration (LfD), but often suffer from high inference cost. This paper introduces Temporal Policy, a generative framework based on stochastic interpolants that formulates action generation as a temporally coupled transport problem. By initializing the generative flow at the robot's recent history, we explicitly couple past states to future action sequences. This data-dependent coupling reduces transport cost and produces straight vector fields. We validate Temporal Policy across visuomotor simulation benchmarks and on a physical Barrett WAM 2x 7DoF teleoperation platform. Our approach reduces transport costs by nearly an order of magnitude compared to noise-initialized baselines, achieving a 19.1 ms inference latency on a single NVIDIA RTX 4080. Crucially, these geometric and computational efficiencies are achieved while matching the success rates of state-of-the-art baselines. This simplified transport geometry bypasses the computational bottleneck of independent Gaussian priors, helping enable high-frequency, closed-loop control. The code is publicly available at https://github.com/dmiller12/TemporalPolicy.
Dylan Miller, Martin Jagersand
Department of Computing Science, University of Alberta, Canada
We present Action Agent, a two-stage framework that unifies agentic navigation video generation with flow-constrained diffusion control for multi-embodiment robot navigation. In Stage I, a large language model (LLM) acts as an orchestration module that selects video diffusion models, refines prompts through iterative validation, and accumulates cross-task memory to synthesize physically plausible first-person navigation videos from language and image inputs. This increases video generation success from 35% (single-shot) to 86% across 50 navigation tasks. In Stage II, we introduce FlowDiT, a Flow-Constrained Diffusion Transformer that converts optimized goal videos and language instructions into continuous velocity commands using action-space denoising diffusion. FlowDiT integrates DINOv2 visual features, learned optical flow for ego-motion representation, and CLIP language embeddings for semantic stopping. We pretrain on the RECON outdoor navigation dataset and fine-tune on 203 Unitree G1 humanoid episodes collected in Isaac Sim to calibrate velocity dynamics. A single 43M-parameter checkpoint achieves 73.2% navigation success in simulation and 64.7% task completion on a real Unitree G1 in unseen indoor environments under open-loop execution, while operating at 40--47 Hz. We evaluate Action Agent across three embodiments: a Unitree G1 humanoid (real hardware), a drone, and a wheeled mobile robot (Isaac Sim), demonstrating that decoupling trajectory imagination from execution yields a scalable and embodiment-aware paradigm for language-guided navigation.
Jeffrin Sam, Nguyen Khang, Yara Mahmoud +2
Intelligent Space Robotics Laboratory, Skoltech, Bolshoy Boulevard, 30, Moscow 121205, Russia