PathLift | CoRL 2026

Overview

Generalist visual-navigation policies tell a robot which direction to go but not how far. A separate lifting step is needed at deployment to convert their unitless trajectory into a metric path that a controller can track. Existing scalar lifters apply uniform scaling that misses per-waypoint corrections, while in-policy approaches that inject monocular depth inherit prediction errors near obstacles and at close range.

We present PathLift, a per-waypoint depth-conditioned adapter for pretrained visual-navigation policies. It consumes an explicit, sensor-grade depth source (RGB-D, stereo, or projected LiDAR) and fits per-waypoint magnitude and lateral residuals to a canonicalized prior. PathLift improves closed-loop navigation across visual-navigation policies in simulation and on a real wheeled robot with stereo and LiDAR depth. It produces a depth-feasible metric path in a single feed-forward pass, without test-time optimization.

Real-robot examples comparing a constant-scale trajectory with a PathLift trajectory around obstacles — Real-robot examples. A constant-scale lift leads into obstacles; PathLift uses depth to adjust the trajectory. Each example shows the camera image and corresponding bird's-eye view.

Method

The adapter receives a canonicalized waypoint sequence and a depth frame projected into the policy camera frame. At each waypoint, it predicts a magnitude and a lateral residual. Training targets are selected from depth-feasible, intent-preserving candidate paths using the expert trajectory. At deployment, the network directly predicts the metric path.

PathLift pipeline: a visual-navigation policy and depth observation are processed to produce a metric trajectory — PathLift inference pipeline. The policy path and reprojected depth are inputs to PathLiftNet; the output is a metric trajectory in body coordinates.

Results

Closed-loop simulation on HM3D

We evaluate 2,000 trials across 50 unseen scenes. The same PathLift adapter is used with all four navigation priors.

Visual-navigation prior	Prior only SR (%)	With PathLift SR (%)	Relative gain	Adapter (ms)
GNM	17.40	25.60	+47.1%	5.3
ViNT	25.00	33.70	+34.8%	5.2
NoMaD	17.45	29.80	+70.8%	5.3
FlowNav	16.00	28.45	+77.8%	5.3

Success rate and adapter inference time. Latency is measured on an NVIDIA RTX 4090.

Image-goal navigation on a real robot

Across three indoor corridors, stereo-depth PathLift achieves a mean success rate of 86.7%, compared with 48.9% for the fixed-scale baselines. With projected LiDAR, aggregated success is 80.0%.

Prior	Fixed scale SR (%)	PathLift + stereo SR (%)
GNM	46.67	86.67
ViNT	66.67	93.33
NoMaD	33.33	80.00

15 trials per method across environments E1–E3.

Composition with local planners

On an 800-trial HM3D subset, PathLift followed by MPPI achieves 38.00% success and 34.57% SPL, compared with 34.12% success and 29.11% SPL for MPPI alone.

Method	SR (%)	SPL (%)	Stage time (ms)
PathLift	29.62	28.20	5.6
MPPI	34.12	29.11	21.3
PathLift → MPPI	38.00	34.57	26.9

Acknowledgments

This work was supported by Hyundai Motor Company and Kia, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) [No. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)].

Reference

Gwanjun Shin, Hyeongseok Suh, Hongmo Jung, and Jongwoo Lim. “PathLift: Depth-Conditioned Metric Lifting for Visual-Navigation Priors.” In Proceedings of the 10th Conference on Robot Learning (CoRL), 2026.