The fast 1-frame bundle dropped the future video tokens that policy mode is trained with and ignored the image. The full bundle follows the image (NVIDIA AV clip: -89 deg, mirrored +90 deg). Rendered frames are JPEG-compressed (q95) before encoding; without compression artifacts the model plans as if stopped.
MBRA docs: goal pose and waypoints are (x, y, cos, sin); spacing 0.8
The wrappers' rectification warp assumes a 512x288 source frame. Callers that pre-resize to the old 64x85 / 96x96 value fed the models a mostly black image.
