Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
From Imitation to Alignment: Human-Preference Flow Policies for Long-Horizon Sidewalk Navigation
TL;DR
Long-horizon sidewalk navigation for autonomous systems faces challenges like unpredictable terrains and social compliance. FlowPilot, a new navigation policy, uses anchored flow matching and human-in-the-loop preference learning to enhance performance.
✦ Why It Matters
Engineers can leverage FlowPilot's techniques to enhance navigation systems in complex, pedestrian-rich environments.
Key Takeaways
- Implement anchored flow matching for diverse behavior representation.
- Utilize human-in-the-loop methods for policy refinement.
- Focus on social compliance in pedestrian environments.
- Test navigation systems in both simulations and real-world scenarios.
Full Summary
Autonomous navigation on sidewalks is crucial for applications like robotic food delivery and assistive devices, but it presents unique challenges compared to road driving, such as navigating unpredictable terrains and interacting with pedestrians. FlowPilot is introduced as a mapless navigation policy that utilizes a monocular RGB camera for efficient long-horizon navigation.
The methodology includes anchored flow matching for pre-training on diverse robot fleet data, capturing complex sidewalk behaviors. To improve the policy further, a human-in-the-loop preference learning scheme is implemented, allowing for fine-tuning based on limited human feedback.
Extensive simulations showed FlowPilot achieving a 42% success rate and 66% route completion, while the enhanced version, FlowPilot-HP, demonstrated improved robustness and social compliance, reducing incident rates by 40% and 52% respectively. These findings suggest significant advancements in the field of robotic navigation, particularly in human-centric environments.