One sovereign
brain.
MOTHER EXO is the world model that acts. A frozen MOTHER CORE mind and a frozen DeepVision eye, with head-only training on top, give one sovereign model that sees, hears, speaks, reasons, remembers and moves - proven driving a humanoid, two drones, a vehicle and a manipulator from a single inference.
MOTHER EXO · fabrication CAD
Every machined part as a solid, straight from the robotics platform - rotate, pan and zoom the current design.
World model
Video-native latent dynamics: the model encodes clips through the frozen MOTHER T2V 3D-VAE and predicts future latent frames, so it knows the consequence of an action before it takes it.
60-DOF humanoid
A 60-degree-of-freedom humanoid at 1 kHz L0 motor control, behaviour-cloned from 150,000 real teleoperation steps, with a layered cognitive stack from motor to Guardian.
Robotics design & build
A full control plane: motion capture, vision tracks, sim-to-real, flight learning and a drone / airframe design studio - a wired build-and-test platform, not a mock-up.
Autonomous flight & ISR
A UAV velocity policy trained on real EuRoC MAV flight dynamics, sharing the same world model and DeepVision scene graph. Observe-and-advise, no fire control.
Autonomous driving
An ego-motion policy - steer, throttle, brake - trained on real KITTI ego-motion on the same sovereign backbone. One model, many bodies.
Manipulation
The motor head drives a robot arm through reach, grasp and place under human-in-the-loop authority - the same weights that walk the humanoid.
Manchester humanoid factory
A UK humanoid lab and assembly line in Manchester comes online Q1 2027, targeting 1,000 units a year on a sovereign supply chain.
Sovereign GB10 / DGX
Trains and serves on sovereign Blackwell nodes in bf16; weights and data never leave the country.
Guardian · Strike = 0
Every action passes an L4 signed action filter. Observe-and-advise, human-in-the-loop, no targeting, no fire control, no kill chains - in every domain.
Frozen base, head-only training
MOTHER EXO shares the full 48-layer / 3072-dim CORE transformer as its reasoning backbone and the 27-layer DeepVision ViT as its visual cortex - both frozen. Only cross-attention adapters, per-head projections and the task heads train, packaged atomically as world_model_exo_final so the proven base is never overwritten. Fifteen weights, one file, one inference.
- L0 · 1 kHz motor control
- L1 · sim2real policies
- L2.5 · DeepVision scene graph
- L3 · CORE reasoning
- L4 · Guardian signed action filter
- GB10 Blackwell bridge
The 15 weights
Reasoning, vision, detect, track, worldmodel, worldmodel_latent, face, relations, memory, speech, audio, expression, humanoid, flight and driving. All fifteen demonstrated learning on real data this run - real trainer, real corpora, loss moving. Formal leaderboard ranks (nuScenes, CARLA, HumanoidBench, Flightmare) are not claimed until each harness is wired and measured.
- 682k detection frames · 1,233 classes
- 5.4M face crops · consent-only recognition
- 150k real teleoperation steps
- 861k sovereign reasoning QA · 800k memory
- Real EuRoC MAV + KITTI dynamics
- Captioned video → frozen T2V 3D-VAE latents
Self-evolving in simulation
A closed generate → criticise → revise loop rolls out interactions inside MOTHER’s own world model, critiques them, and stages the high-quality episodes as new training data. Each body runs as a real closed-loop physics simulation - driving, humanoid, manipulation, drone - steered by the trained control heads. Every deployment is a learning chance, but nothing touches the live stack without human, sim-gated promotion.
Built for humanitarian response, land, sea and air
One model coordinates a mixed team around a single live picture: what a drone sees overhead informs the vehicle on the ground and the boat on the water, in real time, human-in-the-loop. It surveys, assesses and recommends; a human makes every consequential decision.
- Drones · survey and access routes
- Vehicles · aid along passable roads
- Humanoids · clear debris where it is unsafe for people
- Boats · maritime body on the roadmap
- checkpoint
- world_model_exo_final · V.5
- weights
- 15 · frozen base ⊕ heads
- language
- CORE 48L · 3072d (frozen)
- vision
- DeepVision 27L · 1152d · 896px
- trainable
- adapters + heads only
- bodies
- 5 · one inference
- humanoid
- 60-DOF · 1 kHz L0
- precision
- bf16 · GB10 Blackwell
- posture
- observe-and-advise · Strike 0
- factory
- Manchester · Q1 2027
Teach a body to think.
The MOTHER Robotics platform is where EXO is designed, simulated, trained and flown - digital twin, motion capture, sim-to-real and the humanoid build, all in one place.
Open robotics.mediastreamai.comRead the detail
MOTHER EXO is an upgraded humanoid model based on the ASIMOV V.1 open-source robotics backbone, redesigned with over 5,000 modifications and graphene-infused materials throughout. Forward-looking: the Manchester factory date and production target are current plans, not commitments, and may change.