CrossTracer makes robots navigate like humans across bodies
New VLA model adapts navigation for wheeled and legged robots with 28% better performance.
Researchers from Sun Yat-sen University and the University of Hong Kong have developed CrossTracer, a novel vision-language-action (VLA) model framework designed to solve a critical challenge in robotics: cross-embodiment navigation. The team, led by Yao Wang, notes that while VLA models excel at semantic navigation, they often fail to account for embodiment-specific constraints—what works for a wheeled robot may be physically impossible for a legged one.
CrossTracer introduces a two-stage approach: first, a Vision-Language Trace Proposer (VL-Tracer) generates an initial navigation path from egocentric observations and language goals, then a CE-Adapter refines this trace using robot-specific traversability cues. Unlike traditional methods requiring manual annotations, the team trained the refinement module using CE-RRT*, which converts panoptic segmentation into robot-conditioned cost maps. On the NaviTrace benchmark, CrossTracer achieved 45.68 points, outperforming Google's Gemini 2.5 Pro (35.67) by 10.01 points—a 28.1% relative improvement. Real-world deployments on both wheeled and legged robots further demonstrated superior navigation success and execution efficiency.
- CrossTracer uses a VLA model (Vision-Language Trace Proposer) to generate initial navigation traces from language goals and visual inputs.
- CE-Adapter refines traces using robot-specific traversability data, achieving 28.1% better performance than Google's Gemini 2.5 Pro on NaviTrace (45.68 vs. 35.67).
- The system eliminates manual annotation by using CE-RRT* to generate training data from panoptic segmentation, enabling real-world deployment on wheeled and legged robots.
Why It Matters
CrossTracer bridges the gap between semantic navigation and physical constraints, unlocking more adaptable and efficient multi-robot systems.