Technical Summary: LiftNav: Path Planning via Semantic Lifting in TSDF-Guided Gaussian Splatting

Problem Statement

Autonomous robots operating in unknown indoor environments face a fundamental trade-off between geometric precision and semantic understanding. Classical representations like Truncated Signed Distance Functions (TSDF) provide the reliable geometry required for safe collision avoidance but lack object-level semantics. Conversely, photorealistic methods like Gaussian Splatting (GS) offer rich appearance and efficient modeling but suffer from "soft" geometry, which limits their utility for precise obstacle avoidance. Furthermore, embedding semantics directly into GS representations often incurs prohibitive computational costs and training times. There is a need for a navigation framework that combines the geometric rigor of TSDF with the semantic flexibility of GS, enabling efficient, safe, and goal-oriented navigation without requiring dense 3D semantic embeddings.

Methodology

The authors propose LiftNav, a hybrid navigation framework built upon the GSFusion dual-map architecture, which maintains both a Gaussian Splatting map and a TSDF geometry map. The system operates through a real-time pipeline comprising three core components:

1. Semantic Mapping and Lifting

Instead of embedding semantics into the 3D map, LiftNav defers semantic processing to a lightweight 2D stage:

• 2D Inference: A YOLO-based detection model (specifically YOLOv8n-seg) processes RGB frames to identify objects and generate bounding boxes with class labels.
• Spatio-Temporal Lifting: Detected 2D pixels are back-projected into 3D world coordinates using depth values sampled from the GSFusion fused depth map and the current camera pose. To mitigate boundary depth outliers, the system calculates the median 3D position of the masked region.
• Spatial Consolidation: The accumulated 3D points are clustered using DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to form distinct object instances. This step filters noise and assigns unique IDs and centroids to semantic targets, creating waypoints for the planner.

2. Continuous-Time Trajectory Optimization

The path planner generates trajectories directly on the explicit TSDF geometry rather than implicit radiance fields:

• Parameterization: Trajectories are represented as uniform B-Splines (degree $p=8$), ensuring $C^{p-1}$ continuity for smooth velocity, acceleration, and jerk profiles.
• Initialization: A hierarchical initialization strategy is employed. The environment is discretized into an occupancy grid with a safety inflation layer ($1.5 \times r_{robot}$). A 3D A* search finds a collision-free "skeleton" path between projected safe voxels, which seeds the initial B-Spline control points.
• Optimization: The planner solves a constrained optimization problem using the Adam optimizer to minimize a weighted objective function $J(Q)$:
  $$J(Q) = w_{coll}L_{coll} + w_{dist}L_{dist} + w_{acc}L_{acc} + w_{jerk}L_{jerk}$$
  • Collision Avoidance ($L_{coll}$): A novel hinge-loss-based collision penalty is introduced. It combines a primary hinge loss (creating a potential field that aggressively pushes the path away from surfaces within the safety margin) and a secondary exponential term (maintaining a non-zero gradient outside the margin to ensure continuous optimization).
  • Efficiency and Smoothness: The remaining terms penalize path length, acceleration, and jerk to produce graceful motion profiles.
• Target Handling: To address the "Semantic Trap" where target centroids lie inside physical objects (e.g., a TV screen), the final optimized spline is truncated at both ends by the safety margin distance. The endpoints are then linearly connected to the exact start and goal coordinates. This ensures the robot navigates to the surface of the target safely, rather than attempting to collide with the object's interior.

Key Contributions

The paper outlines the following specific contributions:

1. Semantic Lifting for TSDF-based GS: A method to localize semantic targets in 3D by lifting 2D detections onto a TSDF-guided Gaussian Splatting map without dense 3D semantic embeddings.
2. Robust Semantic Navigation Pipeline: A lightweight pipeline integrating YOLO detection, DBSCAN clustering, and TSDF-based planning.
3. Hinge-Loss Collision Penalty: A custom loss function that improves trajectory smoothness and safety by creating a robust potential field for collision avoidance.
4. Evaluation: A comprehensive benchmark demonstrating the framework's feasibility and efficiency compared to state-of-the-art radiance field baselines.

Results

The authors evaluated LiftNav against SplatNav (a baseline using Neural Radiance Fields and convex corridors) on the Replica dataset across two scenarios:

Scenario A: Discretized Goal Benchmarking

• Feasibility: LiftNav achieved a 100% feasibility rate across all scenes, whereas SplatNav failed in 20% of cases (specifically office3).
• Efficiency: LiftNav generated significantly shorter paths (Average Line-of-Sight Length: 3.87m vs. 5.73m for SplatNav).
• Trajectory Quality: LiftNav produced smoother trajectories with a much lower Maximum Jerk (0.15 m/s³ vs. 0.72 m/s³).
• Trade-off: LiftNav had a slightly higher computation time (0.36s vs. 0.25s) due to the continuous optimization, but this was offset by superior path quality.

Scenario B: Semantic Instance Targeting

• Feasibility: LiftNav maintained a 100% feasibility rate for semantic targets. In contrast, SplatNav's feasibility dropped drastically (e.g., to 44% in office0 and 41% in office3) because its convex corridor generation could not mathematically generate a path into a solid surface boundary.
• Safety vs. Success: While LiftNav achieved 100% feasibility, the "strict" success rate (avoiding all geometric collisions with the ground truth mesh) dropped to 79%. This is because the planner intentionally routes the spline to the safety margin boundary to reach the target centroid, resulting in technical collisions at the very end of the trajectory. However, the authors argue this is a necessary trade-off for reaching semantic goals in a safety-constrained manner.
• Performance: LiftNav consistently produced shorter paths (2.96m vs. 3.88m) and significantly smoother motion (Max Jerk 0.11 m/s³ vs. 0.62 m/s³).

Significance and Claims

The paper claims that LiftNav successfully bridges the gap between geometric safety and semantic flexibility. Its primary significance lies in:

• Resolving the "Semantic Trap": Unlike convex corridor methods (SFC) that fail when goals are flush with obstacles, LiftNav's gradient-based optimization with hinge-loss penalties can navigate to semantic targets even in topologically constrained scenarios.
• Efficiency without Dense Embeddings: By decoupling semantic inference from the 3D map, the system avoids the computational overhead of training semantic Gaussians while still enabling object-level navigation.
• Robustness: The framework demonstrates superior robustness in complex indoor environments where traditional radiance field planners fail to find valid trajectories.

The authors modestly acknowledge limitations, noting that the "boundary violations" (collisions at the target) are a result of the semantic lifting strategy placing centroids inside objects, a design choice prioritizing target orientation over final-moment safety. They also note that current evaluation is limited to the Replica dataset, with future work planned for more challenging and outdoor scenarios.