Eximius Labs

Eximius Labs builds open models that give machines senses. Touch, heat, motion and sound become things you can search in plain language, on hardware you own.

The method is the same every time: freeze a strong multimodal base, attach a sense from the outside, and keep every existing embedding bit-for-bit valid. Adding a sensor is a small trained head and an afternoon of compute, not a new foundation model. One shared space spans text, image, video, audio, thermal, motion and touch.

Technical report: https://arxiv.org/abs/2607.18666. On what a training corpus actually teaches an audio embedding: https://arxiv.org/abs/2608.01560. Try a real robot's memory in the live playground: https://www.eximiuslabs.com/playground.

The senses

Each pack is separately loadable and leaves the base model untouched.

• Ember — thermal and infrared, as a loadable sense pack.
• Tremor — body-worn and robot motion (IMU and accelerometer), read as language, with a Unitree-G1 head.
• Tactus — touch (32x32 pressure/taxel arrays from FSR gloves, e-skins, and robot hands), as a loadable sense pack; matches, and at best exceeds, the STAG (Nature 2019) supervised baseline while remaining open-vocabulary.
• Tactus Mat — the same tactile pack trained for a 64x32 body pressure mat (bed, seat, wheelchair, insole): 17 in-bed postures answered as text queries, 0.957 top-1 on held-out subjects.

The base models

The shared space itself: one vector space for cross-modal retrieval, RAG, clustering and semantic search. The senses above attach to these and never modify them.

• fusion-embedding-2 — text, image, video, and audio in one space (connector plus modality-gated adapters).
• fusion-embedding-1 — the connector-only first generation.
• fusion-embedding-2-k3-vision — Kimi K3's vision encoder, projected into the shared space.

Fusion Perception, the perception layer

Dense scene understanding and geometric place recognition on a frozen vision backbone, projected into the Fusion Embedding space. Code: https://github.com/Eximius-Labs/fusion-perception.

Engram, the memory layer

The open cross-modal memory layer for physical AI, built on Fusion Embedding. It indexes a robot's video, audio, and motion into one embedding space on a shared clock and answers questions about it in plain language, including temporal reasoning that retrieval alone cannot do.

pip install engram-robomem

Code: https://github.com/Eximius-Labs/engram · PyPI: https://pypi.org/project/engram-robomem · Playground: https://www.eximiuslabs.com/playground

Efferent, the control layer

The senses read in; Efferent carries commands back out. Any ONNX control policy, any robot, bound by joint name rather than hand-maintained index arrays, with a pre-flight doctor that probes the network to catch permuted joint order, world-vs-body frame errors, missing scales, and dead command dims before the robot moves.

pip install efferent

Code: https://github.com/Eximius-Labs/efferent · PyPI: https://pypi.org/project/efferent

Build with it

• Site: https://www.eximiuslabs.com
• Cookbook: https://github.com/Eximius-Labs/cookbook
• Code: https://github.com/Eximius-Labs
• Paper: https://arxiv.org/abs/2607.18666

Everything is open weight and self-hostable. Core models are research preview. Sensor-pack licenses follow their training data: most packs are CC-BY-NC-4.0, while Tactus Mat is ODC-By 1.0. Code is Apache-2.0.