HALO paper deploys human-preference reward learning on Clearpath Husky

Evidence notes

• The HALO paper introduces an offline reward-learning method that turns human navigation preferences into a vision-based reward model.
• The authors deploy the reward model on a Clearpath Husky across diverse real-world navigation scenarios.
• The paper reports improved success rate, normalized trajectory length and Frechet distance against state-of-the-art vision-based navigation baselines.