NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance · LFM

Identity and contribution

Wenzhe Cai, Jiaqi Peng, and colleagues propose NavDP. The notes use arXiv:2505.08712v3, December 24, 2025; the repository identifies ICRA 2026. The method combines large simulated navigation demonstrations, conditional trajectory diffusion, and privileged geometry-based candidate scoring.

Complete maps/ESDFs provide training labels, while deployment uses local RGB-D and optional 2D goals. “Mapless” describes the deployed policy, not the demonstration engine.

Data engine and embodiment assumptions

A cylinder/differential-drive robot has safety radius 0.25 m, randomized height 0.25–1.25 m, and height-linked camera pitch. Camera height also proxies body clearance; this can fail for mid-body cameras, extended arms, or carried objects.

Geometry is voxelized at five centimeters, sliced by navigation/obstacle heights, converted to ESDF, and planned at twenty-centimeter resolution with A*, clearance refinement, and cubic splines. Published threshold descriptions do not uniquely specify every ground/top/interval rule; actual code is needed to avoid incorrectly blocking under-table space.

Scenes come from 3D-FRONT, HSSD, HM3D, Replica, Gibson, and Matterport3D, rendered with BlenderProc and randomized views/light/textures. The table reports 3,154 scenes, 1,627.1 km, 452 hours, forty million images, and over 200K trajectories. Reported production throughput is specific to its hardware/process.

Conditional generation and scoring

Defaults use eight RGB-history frames, current depth, twenty-four waypoints, and ten diffusion steps. Goal-free behavior zeros the goal embedding. The interface is:

\tau^{(j)}\sim p_\theta(\tau\mid o_t,g_t), \qquad j^*=\arg\max_jV_\theta(o_t,\tau^{(j)}).

The scorer is a supervised geometric evaluator, not necessarily a TD-learned long-term RL critic or calibrated collision probability.

Depth Anything RGB features and a trained depth ViT are compressed into a short sequence; valid depth is 0.1–5 m. History helps when an obstacle leaves view before the robot’s rear clears it, but is not long-term mapping. Shared Transformer bodies support both noise prediction and trajectory scoring. Rotated candidate paths create collision negatives.

Formula boundaries

The stored scoring label sums ESDF differences plus near-obstacle indicators. It does not specify all coefficient signs/values. Under a higher-is-better interpretation, the near-obstacle term needs a penalty direction, which must be verified rather than assumed. The difference sum telescopes to an endpoint difference and cannot alone summarize intermediate risk.

The printed supervised scoring loss is a signed residual without an explicit square/absolute value. Direct minimization would not be a proper regression objective. The notes flag this exact PDF wording instead of silently replacing it with MSE; reproduction should inspect the implementation. Standard DDPM equations in the notes are explanatory expansions of the paper’s compact notation.

Joint training uses batch 2,048, learning rate $10^{-4}$ , and 32 A100s for 24 hours, approximately 768 device-hours. Reusing released weights does not require repeating that budget.

Evidence and ablations

Evaluation	Comparator	NavDP
Simulation PointGoal success	ViPlanner 60.9%	67.2%
Simulation PointGoal SPL	58.6	62.6
Real PointGoal success	53.3%	76.7%
Simulation goal-free time to collision	NoMaD 36.6 s	106.2 s
Simulation exploration area	85.7	274.1
Real goal-free time to collision	29.3 s	112.9 s

Simulation includes 2,000 PointGoal and 1,000 NoGoal episodes. Real quantified comparisons use limited trials across Turtlebot4, Go2, and G1. Collision-free duration is higher-is-better, unlike time to goal. The table-derived exploration-area ratio is about 3.20, slightly different from the rounded narrative.

Removing depth reduces home/commercial success 60.3/74.1%→47.8/66.1%; removing RGB gives 53.9/70.3%; single-frame RGB gives 56.9/72.0%. Random candidate selection reduces it to 53.1/65.6%. Training only low-height embodiments reduces a tall R1’s under-table-scene performance from 90% to 20%. These support the interfaces and data coverage, not arbitrary embodiment transfer guarantees.

Limits and project use

Static planned data dominate training. Pedestrian demos do not establish learned yielding norms, comfort, group behavior, or dynamic safety. Local history cannot solve full exploration memory; natural-language goals and efficient post-training remain directions.

Project proposal: keep perception, generation, scoring, and tracking visible as separate interfaces. Evaluate static contacts, dynamic contacts, proximity, plan timing, and sensor age. A mecanum platform still needs a compatible waypoint-to-velocity controller. Bind camera/depth units, history gaps, goals, robot clearance, actual scoring loss, and repository commit. Later HTTP/asynchronous/MPC infrastructure affects measured behavior and should be recorded separately. No experiments were rerun in the source export.

Source references

The source versions and checks above are retained from the uploaded PaperRead notes. This English edition edits their presentation; it does not report new experiments.