AI

CrossTracer: One Semantic Path, Residual Fixes per Robot

Shar Hendrix 4 min read

A path that makes sense for a dog-like robot can strand a wheeled Scout at the bottom of the stairs. CrossTracer (arXiv:2608.06688, Aug 7, 2026) treats that as the whole problem: keep one semantic pixel-space plan, then add embodiment-specific residuals.

Authors from Peng Cheng Laboratory, SUSTech, and collaborators evaluate on NaviTrace and deploy on wheeled and legged platforms. Project site: lilduckkk.github.io/CrossTracer-Nav.

Legged and wheeled robots in a lobby with different pixel-space navigation traces
Same instruction, different feasible traces: legged robots take stairs; wheeled SCOUTs detour. Source: arXiv:2608.06688.

How it works

  1. VL-Tracer — adapts a pretrained VLA (OmniVLA-style) to propose an initial 8-point normalized image-plane trace from RGB + language and/or pixel goal. Embodiment is not injected here.
  2. CE-Adapter — predicts residual corrections using robot identity (embeddings + FiLM), visual features, and cross-attention from the initial trace.
  3. CE-RRT* — training-time labeler only: Mask2Former panoptic maps → robot-conditioned cost maps → RRT* pixel traces, so the adapter does not need hand-drawn embodiment paths.

At inference, no segmentation planner is required—just observation, goal, and robot type.

CrossTracer two-stage architecture diagram
Stage 1: embodiment-agnostic VLA trace. Stage 2: CE-Adapter residual refinement. Source: arXiv:2608.06688.

Scoreboard

On NaviTrace (official total score; higher is better):

ModelTotal score
CrossTracer-8B45.68
Gemini-2.5-Pro35.67
Robobrain-2.5-8B27.96
CrossTracer w/o CE-Adapter22.56

That is a +10.01 absolute gain over Gemini-2.5-Pro (~28.1% relative). Ablating CE-Adapter drops the total by 23.12 points—especially on accessibility, social norms, and stationary obstacles.

With an extra goal-pose input, CrossTracer reaches 63.91. Embodiment scores stay balanced (bicycle / human / legged / wheeled all ~42–46).

Real-world tests use Jetson Orin on-robot, RTX 4090 offboard inference over Wi-Fi, comparing against OmniVLA under matched tasks; the paper reports improved success and paths closer to human expert references.

A Human’s Take

Pixel-space residuals are a clean interface: language understands “go upstairs to the blue door,” kinematics decide whether that is legal for this chassis. Beating a giant generalist VLM on an embodiment-sensitive benchmark with an 8B stack is a fun receipt. I’d love open weights plus outdoor mud next—not only marble lobbies.

Sources