SEAOTTER: New compression framework beats AVIF with 7x faster encoding

Robotics systems capture high-resolution visual data cheaply, but limited bandwidth and on-device compute prevent full use with traditional codecs like JPEG/MPEG. Newer codecs like AV1/AVIF improve rate-distortion but require custom ASICs for practical encoding, making them unsuitable for resource-constrained robots. Recent asymmetric autoencoders offer high quality under extreme constraints but add prohibitive decoding costs and use bespoke formats ignoring decades of JPEG infrastructure.

SEAOTTER solves this by combining the compactness of a learned latent representation with the broad usability of a standard JPEG file. The key innovation is a learnable JPEG color and quantization transform that avoids performance degradation from naive transcoding. At a 200:1 compression ratio and compared to AVIF, SEAOTTER delivers 7x faster encoding, 3.5x faster decoding, and 8% higher ImageNet top-1 accuracy. The framework supports both general-purpose and task-aware pipelines, enabling cloud robotics to transmit high-quality video without specialized hardware or abandoning existing tools.