A quadruped robot completed 25 of 30 image-goal navigation trials with a new route-scoring model, compared with 13 of 30 for its base navigation policy. The model predicts how candidate routes change a scene in a compact visual representation instead of generating a future video for every route. Researchers from Nanjing University, Imperial College London, Sony AI and Beijing University of Posts and Telecommunications described the system in a paper posted Oct. 8.
A visual navigation policy can suggest several paths from the robot’s current camera view to an image of its destination. A world model can help choose among those paths by estimating what each one would lead to, but systems that render future camera images for every candidate can be slow. LiteNWM was designed to compare routes without reconstructing those future images.
The system receives four recent camera frames and a goal image. Its base policy, NoMaD, proposes 16 candidate routes alongside its current route. LiteNWM predicts the future visual features associated with each route at four short horizons, then scores those predictions against the goal and selects a path. Because the current view is encoded once and reused across candidates, the evaluator can compare several predicted futures without producing a separate video for each one.
Across three navigation datasets, LiteNWM lowered average path-position error by 17.56% compared with NoMaD plus NWM-XL, which generates future images to rescore the same route candidates. Both systems started with the same candidate paths. The paper measures error as root-mean-square position error across four predicted waypoints, scaled by the typical waypoint spacing in each dataset; each dataset contributes equally to the overall result. The comparison used four-step predictions, although NWM was originally designed for eight-step rollouts, which may disadvantage it at the shorter horizon. The offline study used 100 query states in each dataset and four proposal seeds per state, for 1,200 state-seed evaluations. In a separate end-to-end timing test on an RTX 5090, full LiteNWM took 0.691 seconds per decision state, compared with 88.449 seconds for the same NoMaD-plus-NWM-XL setup.
The researchers also tested LiteNWM on a Unitree Go2 EDU quadruped in three scenes: an open indoor area, a cluttered indoor route and an outdoor courtyard. Each method ran 10 trials per scene. LiteNWM reached the goal in 25 of 30 trials; NoMaD alone did so in 13. Success required entering a calibrated one-meter radius with the target visible in a goal-consistent view and meeting a 60-degree yaw-error limit. The paper does not specify the direction used to measure that yaw error. The valid state had to last at least one second or end in a normal stop; collisions, corrective interventions and abnormal termination ended a trial.
The timing results cover different setups. The RTX 5090 comparison was against the image-generating reranker; on the Go2 EDU Jetson Orin, LiteNWM averaged 2.228 seconds per completed planning call, versus 0.651 seconds for NoMaD alone. The physical result covers 30 trials per method on one robot across three scenes, and does not establish performance on other platforms or environments.
Sources
- LiteNWM preprint on arXiv
- Full paper, evaluation protocol and results
- Illustrative archival photograph, not the robot used in the study: Sgt. Amber Edwards / U.S. Army, Unitree Go1 at Romanian Land Forces Day 2023, public domain in the United States
