Item

Vision-and-Language Navigation on a Quadruped Robot: Evaluating NaVid with Temporal Visual History Ablation and Sim-to-Real Deployment

Alameri, Sumaya Abdulrahman Abdulla
Citations
Altmetric:
Supervisor
Department
Robotics
Embargo End Date
2027-05-01
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Vision-and-Language Navigation (VLN) enables embodied agents to follow natural language instructions through indoor environments using visual observations. While significant progress has been made in simulated settings, physically deploying VLN models on real robotic platforms remains rare, particularly on legged systems. This thesis presents two contributions. First, a systematic ablation study is conducted on the temporal visual history mechanism of NaVid, a video-based vision-language model for Vision-and-Language Navigation in Continuous Environments (VLN-CE). Five history conditions are evaluated on the R2R val unseen benchmark: no history, 3 frames, 5 frames, 20 frames, and full history. Results on a common 617-episode subset show that success rate increases monotonically with history length from 9.4 percent to 39.1 percent, while SPL exhibits a non-monotonic pattern with a threshold effect between 5 and 20 frames and a marked improvement between the 20-frame and full history conditions. Second, NaVid is deployed on a Unitree Go2 quadruped robot, demonstrating the feasibility of executing a continuous-control VLN pipeline on a legged platform using only monocular RGB observations, without depth sensors, architectural modifications, or retraining. The results demonstrate that NaVid's navigation performance is strongly sensitive to the quantity of temporal visual context provided at inference time, with both navigation success and path efficiency improving substantially as history length increases. Further, the deployment results demonstrate that a simulation-trained VLN model can produce executable navigation behaviour on a physical quadruped platform under controlled indoor conditions.
Citation
Alameri, Sumaya Abdulrahman Abdulla, "Vision-and-Language Navigation on a Quadruped Robot: Evaluating NaVid with Temporal Visual History Ablation and Sim-to-Real Deployment," M.S. Thesis, Robotics, MBZUAI, 2026.
Source
Conference
Keywords
Quadruped Robots, Vision-Language Navigation, Robot Navigation, Temporal Visual History, Sim-to-Real Deployment
Subjects
Source
Publisher
DOI
Full-text link