Aether for robotics.
The world model robots train in.
The embodied-AI labs build the bodies and the policies. SimOS builds the world those policies learn in — forward-rolled contact dynamics, not scraped text. Twelve specialist agents — world model, contact solver, deformables, synthetic perception, imitation, policy training, manipulation, locomotion, fleet learning and safety verification — under one model, with a calibrated reality gap and an honest fidelity tier on every rollout.
- Agents
- 12
- Path
- Task → Deployed policy
- Classes
- Arm · humanoid · mobile · AV
- Tiers
- Reduced → RT → high-fidelity → FEM
Demonstration to deployment, under one model.
The incumbent robotics stack is a pile of disconnected sim rigs, teleop tooling and one-off transfer scripts. SimOS is one world model and one project file — built so the physics the policy trains in is the same physics the safety case is argued from, without a single re-implemented simulator in between.
Forward world model
SimOS rolls a system forward in time — rigid bodies, contacts, actuators, sensors — rather than describing it in language. The substrate every policy is trained and evaluated against.
Contact & rigid-body solver
Stable, differentiable contact: friction cones, restitution, articulated chains and closed kinematic loops at interactive step rates, accurate enough that the policy that works in sim works on hardware.
Deformables & soft-body
Cloth, cable, soft grippers, granular media and fluids — the contact-rich regimes where teleop data is scarce and a hand-built rig falls over.
Sim-to-real transfer
Calibrated reality gap: system identification from a handful of real rollouts, residual-physics correction, and an honest estimate of where the transfer will hold and where it won't.
Domain randomization
Structured randomization over mass, friction, latency, sensor noise and lighting — sampled from physically plausible priors, not arbitrary ranges, so robustness transfers instead of overfitting.
Synthetic perception
Photoreal RGB, depth, segmentation, tactile and proprioception rendered in the loop — perfectly labelled, infinitely re-sampled, and consistent with the same physics the controller sees.
Imitation from demonstrations
Teleop, kinesthetic and video demonstrations replayed, retargeted across embodiments and augmented in sim — turning a few hundred demos into a training distribution.
Policy training harness
RL, imitation and offline learning against the world model — vectorized environments, reward shaping, curriculum and checkpointing, with the model version hashed into every run.
Manipulation & grasping
Dexterous and parallel-jaw grasping, in-hand reorientation, insertion and tool use — generated and physics-filtered, never a trajectory that violates contact or reach.
Locomotion & whole-body
Bipedal, quadruped and wheeled-legged control with whole-body dynamics, terrain randomization and fall recovery — trained where a real fall costs nothing.
Fleet learning & failure memory
Edge-case rollouts return from the field into the world model; every fleet failure becomes a reproducible scenario the next policy is graded against.
Safety envelope & verification
Formal reachability and constraint checks on the trained policy, an honest operational design domain, and a safety case that says “not yet certified” when coverage is thin.
Our AI for robotics.
Most robotics AI is either a simulator or a single narrow policy. Aether is also the brain — a vision-language-action foundation model that perceives, reasons in language and acts, and generalizes across embodiments. The robot foundation model, on the same model that builds the world it trained in.
Vision-language-action
One model that sees, reasons in language and outputs motor actions — a VLA foundation model, not a pipeline of bolted-together perception, planning and control.
Perception & scene understanding
3D scene, object, affordance and state estimation from cameras, depth and tactile — the world parsed the way an embodied agent needs it.
Language-grounded planning
A natural-language goal decomposed into a plan and grounded in what the robot can actually see and reach right now.
Dexterous control
Closed-loop motor control for manipulation and locomotion — contact-rich, reactive, and trained in the world model where a failure costs nothing.
Cross-embodiment generalization
One brain across arms, humanoids and mobile platforms; a skill learned on one body transfers to another instead of being rebuilt.
On-robot & edge inference
Distilled to run on the robot at control-loop rates, with the full model in the cloud for the hard calls and the long horizon.
Few-shot skill acquisition
Learn a new task from a handful of demonstrations, not a fresh multi-week training campaign per skill.
Closed-loop reactivity
Re-plan and recover when the world doesn't cooperate — a dropped part, a moved target, a scene that changed mid-task.
The difference: those are policies or VLAs trained on demonstrations and scraped video. Aether is trained on simulation and forward-rolls physics — so the brain that acts and the world it learned in are one model, and the reality gap is reported, not discovered on hardware.
Every rollout knows how real it is.
The governing principle: a screening rollout is never presented as high-fidelity, and high-fidelity is never presented as hardware. The four-tier fidelity ladder makes the reality gap structural and reported — not a promise made on deploy day.
Reduced-order dynamics
Analytic and screening models — point-mass, simplified contact, kinematic reach. Milliseconds. For feasibility and design-space sweeps, always labelled as screening.
Real-time rigid-body
Interactive contact + articulation for massively parallel policy rollout. The tier most training runs live in — fast enough for billions of environment steps.
High-fidelity contact
Soft-body, granular and fluid coupling tuned against measured hardware. The accurate physics behind a credible sim-to-real claim.
First-principles multiphysics (escalation)
FEM contact, CFD, thermal and structural coupling. A regime we cannot yet resolve at a given fidelity returns the engine it requires — never a fabricated rollout.
Six robot classes, class-specific physics.
Not one generic environment — each class carries its own dynamics, sensors and failure modes.
Arms, parallel-jaw and dexterous hands. Grasp synthesis, in-hand reorientation, contact-rich insertion and tool use with reach and collision validity gates.
Whole-body bipedal control — balance, locomotion, manipulation and fall recovery trained against randomized terrain and perturbation.
AMRs, quadrupeds and wheeled-legged platforms. Navigation, rough-terrain locomotion and dynamic obstacle response under sensor and actuation noise.
Multirotor and fixed-wing — aerodynamics, wind disturbance and contact for inspection, delivery and perching policies.
Closed-loop driving scenarios with traffic, sensor models and rare-event synthesis — the long tail rendered, not waited for.
Tele-operated and assistive systems with soft-tissue deformables, tight tolerances and a conservative, audited operational envelope.
One solver, five regimes.
The world model is the product. We report what it actually does — including the regimes where the reality gap is widest and escalation to first-principles is required.
Rigid-body, soft-body, granular, fluid, aero
Differentiable contact with friction + restitution
Demonstrations retargeted across embodiments
RGB · depth · segmentation · tactile · proprio
Deterministic, hashed, fully replayable
Reachability + operational design domain
One model for the whole rig.
A dozen disconnected sim, data and transfer tools collapse into one agent, one contract, one project file.
The world and the brain.
The robot-foundation-model labs train a policy on demonstrations and video. Aether is trained on simulation, so it's both the world the policy learns in and the brain that acts — one model.
Training cycles, compressed.
What compounds is policies shipped per quarter, not any single rollout being faster.
From the world model to the safety case — each emitting results under one integrity contract, with the model version hashed into every rollout.
Reduced-order → real-time → high-fidelity → first-principles. Every rollout carries its tier, provenance and where the transfer is trusted.
Not a zoo of disconnected sim rigs — one model that forward-rolls the world every embodiment trains, transfers and is graded in.
An AI engineer, not a sim rig.
It plans, picks fidelity, trains the policy, validates against hardware, and writes the safety case — for an expert audience.
- 01
Reads the task and embodiment
Ingests the robot, the task, the scene and the constraints, and decomposes the brief into a typed training graph before a single environment step runs.
- 02
Picks the right fidelity tier
Knows when reduced-order dynamics suffice, when to roll out in real-time contact, and when high-fidelity soft-body or FEM is required — by the transfer the conclusion needs.
- 03
Generates and physics-filters
Trajectory and grasp proposals are disposed of by contact, reach and stability checks. Generative plausibility never overrides a violated constraint.
- 04
Validates against real rollouts
Cross-checks sim behaviour against hardware logs, calibrates the reality gap, and flags out-of-domain conditions instead of silently extrapolating.
- 05
Writes the rationale
Each policy ships with its operational design domain, the randomization it survived, the failure modes it didn't, and the recommended next experiment — with the model-version hash.
- 06
Closes the loop
Field failures return as reproducible scenarios in the world model; the policy that shipped gets graded against the edge cases the fleet just found.
Embodied intelligence that improves with every rollout.
Each field failure returns into the world model as a reproducible scenario, so the model that trained a policy gets better at training the next one — and the fleet that ran it gets safer.
“The model told us which task wasn't ready before the robot did. That's the part we didn't have before.”
Train a policy in the world model.
SimOS forward-rolls the world your robots train in — contact, perception and transfer under one model. Request access to deploy it behind your firewall, or see how it couples to the rest of the platform.
Robotics sits on the same foundation as the rest of Aether — one model, every discipline.