Technical Summary: SanD-Planner

Problem Statement

Robust local navigation in highly cluttered and dynamic environments remains a fundamental challenge for mobile robotics. Traditional modular pipelines (perception-SLAM-planning) often fail in dynamic settings due to latency, error accumulation, and reliance on precise mapping. While end-to-end learning approaches, particularly Imitation Learning (IL), offer a promising alternative, they face significant bottlenecks:

1. Data Scarcity and Efficiency: Current state-of-the-art (SoTA) methods rely heavily on massive-scale expert demonstrations (e.g., hundreds of thousands of trajectories) and extensive computational resources, making them difficult to deploy in data-scarce domains like marine robotics or specialized environments.
2. Representation Limitations: Existing IL planners often predict discrete waypoints or interpolating splines. These representations can suffer from myopic behavior (short planning horizons) or lack robustness to perceptual noise, where errors in distant observations propagate to destabilize immediate actions.
3. Safety Verification: Integrating strict safety constraints into the learning process often requires complex reward engineering or value-function learning, increasing training burden and reducing sample efficiency.

The paper asks whether a robust local planning policy can be learned effectively in a "data desert" (limited expert demonstrations) by leveraging structural geometric priors and decoupling safety verification from policy learning.

Methodology: SanD-Planner

SanD-Planner is a sample-efficient, diffusion-based local planner that maps depth observations directly to smooth, parametric paths within a clamped cubic B-spline space. The system operates via a two-stage "generate-and-select" pipeline:

1. Perception and Condition Encoding

The agent receives a history of four depth images ($O_t$), a relative 3D goal ($g_t$), and a previous motion context ($v^{prev}_t$) representing the heading of the executed trajectory.

• Visual Encoding: Depth frames are processed via a lightweight ResNet-18 backbone to extract spatial feature maps, preserving fine-grained geometric details (e.g., obstacle edges).
• Fusion: Visual tokens, goal tokens, and velocity tokens are concatenated and aggregated by a two-layer Transformer encoder to form a multi-modal context tensor ($C_t$).

2. Trajectory Generation (Diffusion Policy)

Instead of predicting discrete waypoints, the policy predicts a fixed set of eight B-spline control points ($Q_t$).

• Inductive Bias: The B-spline representation inherently enforces $C^2$ continuity (smoothness) and utilizes the local support property. Noise or occlusions in far-field observations affect only distal control points, preventing global path destabilization.
• Diffusion Model: A conditional Denoising Diffusion Probabilistic Model (DDPM) with v-prediction parameterization is trained to model the distribution of expert control points given the context $C_t$.
• Temporal Consistency: To prevent oscillatory behavior in multi-modal environments (e.g., choosing left vs. right of an obstacle), the previous plan's heading direction is explicitly fed back as a conditioning token.

3. Safety-Aware Candidate Selection (Critic Module)

To decouple complex distribution modeling from strict safety verification, the system employs an explicit geometric critic rather than learning a safety value function.

• Generation: The diffusion policy samples $K$ candidate control-point sets, which are mapped to continuous trajectories.
• Selection: An ESDF-based (Euclidean Signed Distance Field) critic evaluates candidates using a cost function $J$ comprising:
  • Discounted Safety Cost: Penalizes proximity to obstacles, with higher weight on near-horizon segments.
  • Path Efficiency: Penalizes path length and deviation from the goal.
• The trajectory with the minimum cost is selected for execution in a receding-horizon manner.

Key Contributions

1. SanD-Planner Architecture: A novel diffusion-based planner operating in a clamped B-spline control-point space. It achieves state-of-the-art performance using only 500 expert trajectories (approx. 0.25% of the data scale used by baselines like NavDP).
2. Representation Study: A systematic investigation demonstrating that B-spline control points outperform discrete waypoints and interpolating cubic splines in imitation learning. The study highlights B-splines' advantages in parameter efficiency, inherent smoothness (as an inductive bias), and representation-level robustness against perceptual noise.
3. Zero-Shot Sim-to-Real Validation: Successful deployment on a Unitree Go2 quadruped robot in diverse real-world environments (narrow passages, cluttered offices, stairs) without any real-world fine-tuning or domain adaptation.
4. Reproducibility: The authors commit to open-sourcing the dataset, pre-trained models, and the complete training/evaluation framework.

Experimental Results

Experiments were conducted on the InternNav Benchmark (ClutteredEnv and InternScenes) and in real-world settings.

• Performance vs. Baselines: Despite training on only 500 episodes, SanD-Planner achieved a 90.1% Success Rate (SR) in ClutteredEnv and 72.0% SR in InternScenes. This outperforms or matches SoTA baselines (iPlanner, ViPlanner, NavDP) that utilized datasets ranging from 30k images to 200k trajectories.
• Sample Efficiency: The model demonstrated high efficiency in data-scarce regimes, achieving a non-trivial 55.5% SR with only 50 trajectories (10% of the dataset). Performance scaled monotonically with data size, though the 500-episode model was sufficient for SoTA results.
• Ablation Studies:
  • Representation: B-splines significantly outperformed waypoints and interpolating splines, particularly in long-horizon scenarios, due to better isolation of far-field uncertainty.
  • Velocity Token: Removing the previous velocity token caused a performance drop (e.g., 86.0% to 93.0% SR in ClutteredEnv), confirming its necessity for maintaining topological consistency and preventing oscillatory steering.
• Real-World Deployment: The system successfully navigated static and dynamic obstacles, narrow mazes, and stairs in zero-shot transfer, operating at ~10Hz.

Significance and Claims

The paper claims that robust local navigation does not necessarily require massive datasets if the learning process is guided by strong structural priors. By shifting the learning objective from predicting raw trajectories to predicting compact B-spline control points, SanD-Planner:

1. Reduces Learning Complexity: Smoothness and safety bounds are inherent to the representation, simplifying the policy's learning task.
2. Decouples Safety: By using an explicit geometric critic, the system avoids the difficulty of learning safety constraints via reward engineering or value functions.
3. Enables Data Efficiency: It demonstrates that high-performance navigation can be achieved with a fraction of the data required by current scaling trends, making advanced navigation more accessible for data-constrained domains.

The authors acknowledge limitations regarding small or specular objects due to depth sensor constraints and suggest future work integrating visual foundation models for improved metric depth estimation. However, the primary contribution remains the demonstration that sample-efficient, robust planning is achievable through structured representation and decoupled safety verification.