Industrial ReMEmbR
A shared spatiotemporal memory system for robot fleets proposed by NVIDIA. It extends the single-robot memory framework ReMEmbR into a multi-robot shared memory pool supporting retrieval by content, location, and time, and reduces server-side computation through an edge-cloud split.
Timeline (last 90 days)
NVIDIA proposed Industrial ReMEmbR, extending the single-robot memory system ReMEmbR into a shared spatiotemporal memory for robot fleets.
Industrial ReMEmbR allows multiple robots to write to the same memory pool, and any robot can retrieve observations from other robots.
Industrial ReMEmbR deploys NVIDIA Jetson on each robot for edge encoding.
Industrial ReMEmbR uses Cosmos-Embed1 to generate embeddings for retrieval, Cosmos3-Edge to generate visual tokens, and Wan2.2 VAE to generate reconstruction latents.
Robot-side encoding in Industrial ReMEmbR does not require uploading raw camera frames, only the compressed representations.
In 450 test decodes of the Industrial ReMEmbR system, the output of the cached-token decoding was exactly identical to that of the unsplit model.
Industrial ReMEmbR splits Cosmos3-Edge along the vision-language boundary: the vision encoder stays on the robot, and the language decoder is placed on a local server.
Industrial ReMEmbR does not generate descriptions during the write phase, and generates scene descriptions only when there is a query need (on-demand generation).
At query time, Industrial ReMEmbR uses a Nemotron agent to retrieve memories by content, location, and time, and issues queries to the language decoder using cached visual tokens.
Industrial ReMEmbR reuses the same set of encoding models to detect people in image patches and zero out the corresponding regions, achieving PII abstraction.
2 more timeline events
All Facts (12)
NVIDIA proposed Industrial ReMEmbR, extending the single-robot memory system ReMEmbR into a shared spatiotemporal memory for robot fleets.
50%UnverifiedIndustrial ReMEmbR allows multiple robots to write to the same memory pool, and any robot can retrieve observations from other robots.
50%UnverifiedIndustrial ReMEmbR deploys NVIDIA Jetson on each robot for edge encoding.
50%UnverifiedIndustrial ReMEmbR uses Cosmos-Embed1 to generate embeddings for retrieval, Cosmos3-Edge to generate visual tokens, and Wan2.2 VAE to generate reconstruction latents.
50%UnverifiedRobot-side encoding in Industrial ReMEmbR does not require uploading raw camera frames, only the compressed representations.
50%UnverifiedIn 450 test decodes of the Industrial ReMEmbR system, the output of the cached-token decoding was exactly identical to that of the unsplit model.
50%UnverifiedIndustrial ReMEmbR splits Cosmos3-Edge along the vision-language boundary: the vision encoder stays on the robot, and the language decoder is placed on a local server.
50%UnverifiedIndustrial ReMEmbR does not generate descriptions during the write phase, and generates scene descriptions only when there is a query need (on-demand generation).
50%UnverifiedAt query time, Industrial ReMEmbR uses a Nemotron agent to retrieve memories by content, location, and time, and issues queries to the language decoder using cached visual tokens.
50%UnverifiedIndustrial ReMEmbR reuses the same set of encoding models to detect people in image patches and zero out the corresponding regions, achieving PII abstraction.
50%UnverifiedOn-demand description generation reduces the normal load, but latency behavior at query peaks and the management of cached visual token lifetimes are engineering details that need verification before deployment.
50%UnverifiedThe 450-decode consistency test only covers output consistency and does not yet address memory drift over long-term operation.
50%