UNIT 01-F / RUNNING
[ HIGH OUTPUT ]
PINN ENGINE DEPLOYED
MISSION READY>_
// Subroutine 001 — Brief
Bring your governing equations and sparse sensor data. Astraea builds optimized physics-informed neural-network surrogates that drop into your MPC, RL, or state-estimation pipeline — with automated architecture search, loss balancing, and convergence diagnostics.
// Request early access
// Or install the core
$pip install astraeaSee use cases →// The problem
Policies trained in simulation fail on real hardware because physics engines use approximated contact, linearized friction, and simplified dynamics. The sim-to-real gap is a physics problem, not a data problem.
MuJoCo, Isaac Sim, and PyBullet relax friction cones, soften contacts, and use penalty-based compliance. A policy that hits 95% in sim drops to 30–60% on real hardware.
Underwater, surgical, nuclear, agricultural — you can’t run 10,000 demonstrations when each deployment costs thousands, or when the environment is lethal.
Physics-informed neural networks solve this — but manually choosing architectures, loss weights, sampling strategies, and debugging convergence takes weeks of expert tuning.
Every PINN for robotics is a one-off research effort.
// How it works
Your governing equations are the input, not the code. Astraea validates the problem, searches architectures and losses automatically, and emits a surrogate ready for your control pipeline.
Provide your ODEs/PDEs — Fossen, Cosserat, Navier-Stokes, or any custom equations. Add boundary and initial conditions.
Automatic well-posedness checking. Classifies equation type. Flags stiffness, discontinuities, and ill-conditioning before wasting GPU hours.
Bayesian optimization over architectures, activation functions, loss weights, and sampling strategies. Self-adaptive training balances residuals automatically.
Export to CasADi, Drake, ROS, ONNX, or TorchScript. Real-time inference at 100+ Hz on GPU. Full diagnostics and error bounds included.
// Use cases
Every use case has the same structure: known governing equations, sparse real-world data, need for an accurate dynamics model. The equations are the variable — the pipeline is the same.
Mν̇ + C(ν)ν + D(ν)ν + g(η) = τFossen 6-DOF dynamics with ocean current estimation. Correct DVL drift, estimate current fields, plan paths through currents from sparse IMU and DVL data.
ρA ∂²r/∂t² = ∂n/∂s + fCosserat rod and continuum mechanics for pneumatic actuators, cable-driven arms, and deformable manipulators. Proven 44,000× speedup enabling MPC at 70 Hz.
σ = C : ε + η ∂ε/∂tDeformable tissue modeling for surgical planning and force estimation. Every anatomy is different — PINNs generalize from sparse force-torque data where simulators fail.
M(q)q̈ + C(q,q̇)q̇ + τ_f(q̇) = τFriction-inclusive robot dynamics for high-precision industrial arms. Physics-informed friction models that transfer across operating conditions.
F = ma + F_contact + F_constraintNuclear decommissioning, deep mining, space exploration. When data collection is dangerous or lethal, physics constraints replace the demonstrations you can’t collect.
∇·σ + ρb = ρüSoft contact with plants, fruit, and soil. Variable terrain, seasonal changes, deformable grasping. Each object is unique — physics generalizes where data memorizes.
// Equation library
Start from validated equation templates with known-good architecture configs, or define any custom ODE/PDE system. Each template ships with tested hyperparameter ranges.
// Open source + commercial
The PINN engine and equation library are open source under Apache 2.0. AutoML, managed compute, and deployment integrations are commercial.
Full PINN engine, equation library, and training loop. Everything you need to build physics-informed dynamics models.
Automated architecture search, managed GPU training, and production deployment integrations for teams shipping real robots.
// Why a tool?
There's a difference between generating code and reliably producing dynamics models that control real robots.
| CAPABILITY | LLM + AGENT | PINN LIB | ASTRAEA |
|---|---|---|---|
| Generate PINN code | ✓ with bugs | ✓ manual | ✓ automatic |
| Architecture search | vibes-based | manual | Bayesian opt |
| Loss weight tuning | trial / error | manual | self-adaptive |
| Well-posedness check | no | no | automatic |
| Convergence guarantee | no | no | validated configs |
| MPC / RL pipeline export | manual wiring | manual wiring | CasADi · Drake · ROS |
| Time to working model | days–weeks | weeks | hours |
| Reproducible | no | partially | fully versioned |
[[[ MISSION READY ]]]
Astraea is in early development. Star the repo, join the Discord, or reach out if you're working on physics-informed dynamics for robotics.
// Or request Astraea Pro early access