Education · Robotics & embodied learning
Robotics & learning lab
A real robot-physics sandbox and machine-learning lab, built for education: families with children, schools, and students of any age, from their first years to decades in. Drive robots, build worlds, write robot brains, and teach them to learn by themselves: by practice, by evolution, or by copying you.
The idea
The real thing, made approachable.
Real physics
A research-grade rigid-body physics engine of the kind robotics labs use, with real bodies, joints, motors and sensors, not a game engine pretending.
Real learning
The same families of algorithms used in research, written to be read: every line commented for someone curious but new.
Built for learning together
Plain language throughout, guided lessons from physics to machine learning, and every claim in a lesson measured on a real run and kept true by a test.
How a brain learns
Three ways to learn, and the tools to know whether it worked.
Practice
Reinforcement learning: the robot tries things, gets points, and slowly does more of what earned them, with parallel worlds on several processor cores.
Evolution
A population of brains competes and the best become parents of the next generation, by genetic algorithms or evolution strategies.
Copying a teacher
Behavior cloning from a hand-written controller or from a person driving, with noise added to the teacher so the student learns to recover, then fine-tuned by practice.
Curricula & races
Chain harder worlds in one job, or train several variants side by side: different seeds, sizes, learning rates or methods.
Measured recipes
Missions too hard to stumble on by chance start from a measured recipe, usually copying the teacher first, with the reason shown.
Practice partners
For games, a new brain first plays a statue, then a wanderer, and moves up to the real opponent as it wins. Exams are always against the real one.
Eyes
Brains can learn from the robot's own camera, rendered identically in training and in play, through a small convolutional network.
Memory
Recurrent brains for missions that need memory, and a mission built to show why: without memory a brain can only guess.
Physics & robots
Eleven robot families, and the physics they live in.
Realistic physics
Gravity, air drag, water with buoyancy and currents, gusting wind, bounciness, and floor patches of ice, mud, sand, rubber and trampoline.
Real-robot imperfections
Motor lag and sensor noise, and randomized friction, mass and starting positions, so brains trained in simulation cope with the real world.
Ground, air and water
A two-wheeled rover, a four-legged walker, a quadrotor with its own flight computer, a cart-pole, a race car with real steering geometry, a truck towing a trailer, a motorboat and a sailboat that tacks upwind.
Arms and humanoids
A four-joint arm with a gripper and wrist camera, and a child-sized humanoid with a balance computer that can learn to stand, recover from shoves and walk.
Classical control as teachers
Hand-written controllers for every robot: proportional and PID control, a rhythmic gait, cascaded flight control, inverse kinematics and pure-pursuit path tracking.
Robots built from a recipe
Creatures whose body and brain are one genome, evolved together.
Worlds, games & arenas
Build a world, give it a mission, and let robots compete.
World builder
Arenas, procedural hills, seeded mazes, stairs and craters, with robots, objects, goals and the mission placed by hand and saved to share.
Missions with explained rewards
Find a goal, collect targets, walk, balance, hover, push, pick and place, park, race, and find by color, each with its reward explained.
Games
Sumo, soccer for one or two a side, tag and cooperative pushing. Senses are relative to the robot, so one brain can play any seat, including against copies of itself.
Tournaments & ratings
Round-robin tournaments with side swaps, standings, head-to-head tables and a rating ladder per game that every match moves.
Replays
Every arena game recorded compactly and replayable in slow motion, frame by frame, from any camera. Any world can save its last half-minute.
Twins, save points & rooms
Branch a live world, physics state and all, into a twin and give it a different brain; rewind to save points; and bring a 3D scan of a real room in as the backdrop.
Robot societies
Robots with minds, bodies with skills.
Minds that choose
Each robot can have a mind that picks from abilities such as going somewhere, following, looking, picking up, delivering and building. The body carries the skill out with real physics.
Fast rules, slower thought
Rule-based minds act instantly; language-model minds running on the same computer plan in the background while reflexes keep the robot busy, so physics never waits.
Swarms
From a small village of thinking robots to hundreds of rule-driven workers led by a few planners, with spoken orders and every robot's reasoning open to inspection.
A skill ladder
Every skill climbs from a hand-written script, to a brain copied from it, to a brain that practiced and can beat its teacher. All take the same exam, and the champion is what robots use.
Brain library
Every brain kept, measured and traceable.
Metrics & time machine
Every run's per-epoch metrics, best and latest checkpoints, and periodic snapshots to compare a brain early and late in its training.
Lineage
Fork a brain, train its children and see the family tree across generations.
Exams & exchange
Exams, notes and renaming, and export and import as a single file.
Parallel jobs
Training jobs run in separate processes, one per core, queued beyond that, and can be stopped at any point keeping what was learned.
A standard interface
Every world is a standard learning environment, so any reinforcement-learning library can train on it, from code or the command line.
To real hardware
Trained brains can be exported for small onboard computers on physical robots.
Learning together
For families with children, students of every age, and the people who teach them.
A lesson curriculum
Hands-on lessons from physics and sensors through control, rewards, reinforcement learning, evolution, cloning and language models, with quizzes and challenges.
Profiles & a grown-ups view
A profile per learner and a dashboard with time on screen, lessons finished, quiz results and trained brains, with break reminders and a PIN.
Play together
Several screens on a home network can join the same world, each driving its own robot.
Everything stays local
It installs like an app, runs on the learner's own computer, and nothing leaves it unless they send it.
Connections
The same ideas, elsewhere in our work.
Market simulation research
Learning inside a model of the world and then facing the real one is the shared problem: a market model for a trading policy, a physics engine for a robot.
League operations platform
The same physics, read from the other side: footage of real play turned into positions, speeds and the flight of the ball.
Agent runtime & orchestration
A society of robot minds and a swarm of software agents share one design question: what each one is allowed to do, and how a refusal is enforced and recorded.
Security assurance platform
A lab built for learners of every age is held to the rules that protect children's data, and checked against them continuously.
Engage
By invitation.
Every engagement begins with a conversation and is scoped before work begins.
Families, schools & educators
At home or in class, with learners of any age, and a hand in shaping the curriculum.
Subject: AccessPartnerships
Education, robotics and hardware partners.
Subject: PartnershipCustom simulation
Simulated robots, worlds and training pipelines for your own problem.
Subject: BuildInvestment
Briefings and a demonstration, under confidentiality.
Subject: Fundingcontact@carbonyl.org