Dan Jacobellis, Neeraja J. Yadwadkar
4 min
Abstract
In robotics systems, vast amounts of visual data are easily captured at high resolution using low-cost, low-power hardware. Yet, limited bandwidth and on-device compute resources prevent full utilization when transmitted via conventional codecs like JPEG/MPEG. Newer codecs, like AV1/AVIF, improve the rate-distortion trade-off, but demand far more resources for encoding, impractical without custom ASICs. Recent asymmetric autoencoders deliver high quality under extreme power and bandwidth constraints, but add prohibitive decoding cost and use bespoke formats that ignore decades of infrastructure built around standards like JPEG. To address these limitations, we introduce a compression framework for cloud robotics based on a Sensor Embedded Autoencoder paired with a One-Time Transcode for Efficient Reconstruction (SEAOTTER). Because the sensor, cloud, and consumer stages face very different power and bandwidth budgets, SEAOTTER combines the compactness of a learned latent with the broad usability of a standard JPEG file. Since naive transcoding degrades performance, we propose a learnable JPEG color and quantization transform that enables increased accuracy for global, dense, and vision-language-based perception. Using SEAOTTER, we train both general-purpose and task-aware transcoding pipelines for a pre-trained, frozen encoder. At a compression ratio of 200:1 and compared to AVIF, we observe 7 times faster encoding, 3.5 times faster decoding, and +8% ImageNet top-1 accuracy, while retaining compatibility with JPEG infrastructure. Our code is available at https://github.com/UT-SysML/seaotter .
Alex: The JPEG Sandwich—I like that name. So it's like pre-packaging the data in a way that survives the translation step without losing the details that matter?
Sam: That is exactly right. And the training process is clever. The system learns to predict how a standard JPEG encoder would handle the data, and it adjusts the shorthand accordingly—essentially learning the exact packaging that results in the smallest file size while keeping the image clear enough for machine vision to work reliably.
Alex: Does it actually perform better than what we have now?
Sam: The evidence suggests it is a meaningful improvement. Compared to modern standards like AVIF, this approach achieved much faster encoding on the device while maintaining higher accuracy for machine vision tasks—and it outputs a file that any standard software can open without modification.
Alex: So it's faster for the robot, better for the AI, and compatible with everything else. That's a significant combination of benefits.
Sam: It is. But the researchers are careful to note it is not a "set it and forget it" system. The cloud-side translator needs to be calibrated to the specific AI task you are running. If the AI model changes, that cloud-side component may need fine-tuning. So you are making a deliberate trade-off: you gain enormous efficiency on the robot, but you accept some additional work on the cloud side when things change.
Alex: So you're essentially locking in the robot's behavior to keep it lightweight, and accepting that the cloud has to adapt when needed.
Sam: Correct. And for many real-world deployments—where the robot's task is fixed but battery life is critical—that is a very practical trade-off to make. The robot stays simple and efficient, and the cloud handles the complexity. That division of responsibility is, in many ways, the core insight of the whole paper.
Alex: It's a useful reminder that good engineering is often about choosing where to put the burden, not eliminating it entirely. Thanks for listening to ResearchPod.