Infer

Compressed camera images as base64 strings or data:image/…;base64 URIs, ordered by camera_order and timestep for trajectory modes. When camera_order is omitted for trajectory modes, the legacy 4 camera x 4 frame order is used. Answer-only VQA accepts variable image counts and preserves input order.

Inference mode: trajectory generation, visual question answering, or both.

Optional camera names matching the image camera-major order. Names may use Alpamayo full camera names or OSS short names such as cross_left and front_wide. Required for trajectory modes when image count is not 16.

Egomotion payload required for trajectory modes. ego_history_xyz must have shape [16, 3]. ego_history_rot must contain exactly 16 quaternion rows [w, x, y, z] or 16 3x3 rotation matrices. The 16-row history is required even for reduced image layouts.

Question for mode=‘answer’ or mode=‘both’.

VLM nucleus sampling top-p.

Optional VLM top-k sampling limit. Use 1 for greedy/top-1 decoding.

Optional per-request sampling seed. When omitted, the container default seed is used. For repeatable K=1 trajectory inference in the same deployment, use a fixed seed with top_k=1 and top_p=1.