hzxbzp/test-models
AnchorReason
Checkpoints fine-tuned on **AnchorReasoning**, a visual-grounding and causal-reasoning annotation layer for long-tail driving built on WOD-E2E.
Each model takes one forward panorama + 4 s of ego motion history + a navigation intent and produces a visually grounded chain of thought: scene context, traffic events, the decision-critical elements with their image locations, attributes, impact ranks and per-element implications, then a rationale, an action plan, and a 5 s future trajectory.
Checkpoints
Input format
Getting this exactly right matters: these are full fine-tunes on a fixed prompt, and a prompt that differs from the training one degrades the output badly.
Image. One 2916 x 1079 forward panorama, stitched from the FRONT_LEFT, FRONT and FRONT_RIGHT cameras of a WOD-E2E frame by reprojecting all three onto a single vehicle-aligned virtual pinhole. The dataset card carries the exact reconstruction code. Images are fed at full resolution (long edge 2916).
User turn.
History (past 4 s, 16 waypoints at 4 Hz, ego frame, meters): [x, y], [x, y], ... (16 points)
Intent: GO_STRAIGHT | GO_LEFT | GO_RIGHT
<instruction>Chat scaffold (Qwen2 family, identical for both backbone families):
<|im_start|>system
{system}<|im_end|>
<|im_start|>user
<|vision_start|>{image_pads}<|vision_end|>{user}<|im_end|>
<|im_start|>assistantPoint coordinates
<point> values follow the backbone's own convention:
Each folder ships a prompt_hash.json recording the exact prompt fingerprint the weights were trained with — check it if you rebuild the prompt yourself.
Output format
stage2 emits the whole chain in a fixed tag order; traj_only emits <traj> alone.
<ego_state>lon=... lat=...</ego_state>
<context>weather=...;daytime=...;visibility=...;scenario=...;road=...</context>
<events>...</events>
<has_objects>yes|no</has_objects>
<obj type=...><point>x,y</point><rank>1</rank>...<implication>...</implication></obj>
<n_objects>N</n_objects>
<reason>...</reason>
<final_plan>...</final_plan>
<motion>lon=... lat=...</motion>
<traj>[x1,y1],[x2,y2],[x3,y3],[x4,y4],[x5,y5]</traj>Base models and licenses
Every checkpoint is a derivative of its backbone and stays bound by that backbone's license. They differ, so this repository carries no single license — check the row for the checkpoint you use.
Gated backbones. The two Cosmos-Reason2 base models are distributed behind a license gate. Before using cosmos-reason2-2b/* or cosmos-reason2-8b/*, accept the NVIDIA Open Model License Agreement on the base model page — Cosmos-Reason2-2B or Cosmos-Reason2-8B — where access is granted automatically once you are signed in and have agreed to the terms. These derivative weights are offered on the same terms; the gate lives on the upstream repository, not here.
The AnchorReasoning annotations used to train these models are released under CC BY-NC 4.0, and the underlying imagery comes from WOD-E2E under the Waymo Open Dataset License. Treat the checkpoints as non-commercial research artifacts.
Intended use and limitations
These are research checkpoints for studying grounded reasoning and planning in long-tail driving. They are not driving software: they run on a single forward panorama with no rear or side view, no LiDAR, no HD map and no temporal context, they predict a 5 s open-loop trajectory rather than closing any control loop, and their reasoning text is generated, not verified. Do not deploy them on a vehicle.
Citation
The accompanying paper is under review; a citation will be added once it is public.
@misc{anchorreasoning,
title = {AnchorReasoning: A Visual Grounding and Causal Reasoning Dataset in Long-Tail Autonomous Driving Scenarios},
year = {2026},
note = {Under review}
}