01
Out-of-the-box generalization
The intelligence is built to work on new tasks and in new environments — not narrowly trained for one application and stranded everywhere else.
Built to give machines broad, real-world dexterity — CALISTA is the physical AI behind general-purpose robots.
One robot intelligence that perceives, reasons, and manipulates — paired with a full-stack platform that trains and validates new skills in simulation, then carries them to real hardware.
›
crack two eggs into the pan and cook them
One brain, many tasks
A single robot intelligence performs a wide spectrum of dexterous, real-world tasks — autonomously, at normal human speed.
Cooking eggs
Force-controlled grasp & crack
Pipetting in a lab
Sub-millimeter precision
Making a smoothie
Multi-step sequencing
Solving a Rubik's cube
In-hand reorientation
Picking up objects
Cluttered, varied shapes
Assembling wire harnesses
Deformable, contact-rich
Playing the piano
Timed, bimanual control
New tasks, out of the box
No bespoke retraining
The robot intelligence
One system that perceives, reasons, and acts.
At the center of CALISTA is a robot intelligence system — effectively the brain of our robots. It lets a machine take in its surroundings, understand an instruction in plain language, reason about how to carry it out, and manipulate objects with precision to see it through.
- Perceive. Fuses vision, depth, touch, and force into a live model of the scene and the objects in it.
- Understand. Maps everyday instructions to goals, constraints, and the steps that satisfy them.
- Reason. Plans a sequence, adapts mid-task as the world changes, and recovers when something slips.
- Manipulate. Closes the loop with dexterous, contact-rich control — gentle on an egg, exact on a connector.
Set up the pan and start two eggs while I prep the rest.
Understood. Localizing the eggs, pan, and spatula, then planning a force-controlled grasp. I'll monitor doneness with the overhead camera and report when they're ready to plate.
plan.grasp_egg + stable
If one sticks, adjust — don't force it.
Got it. Switching to compliant control near the shell; if I sense resistance above threshold I'll re-grasp instead of pushing through.
force.limit 1.4 N
The development platform
Thousands of trials in minutes — matched to reality.
Robots are meant to automate work, yet building them still depends on slow, manual training. Our full-stack platform breaks that bottleneck: it runs thousands of simulated trials in minutes, and the results match real-world performance — so a skill can be developed and validated long before it ever reaches the floor.
- Fast. Massively parallel simulation compresses thousands of attempts into minutes, not weeks of manual data collection.
- Faithful. Simulated outcomes line up with what the hardware actually does, so results transfer instead of surprising you.
- Validated. Every skill is stress-tested across variation before deployment, with metrics you can trust.
4,096 trials · 00:51 elapsed
sim ↔ real Δ 1.8%
Full-stack & general-purpose
Full-stack, and built to generalize.
We don't narrow CALISTA to a single industry. The same system handles household activities, laboratory work, and manual assembly — because we build the whole stack, both the robots and the simulation platform that trains them, toward one goal: working out of the box on unfamiliar tasks and environments.
- One stack, end to end. Hardware, intelligence, and the training platform are designed together, not bolted on.
- General-purpose. Versatility across many tasks and environments — not a machine rebuilt for every job.
- Out of the box. The aim is generalization to new tasks and settings the robot was never explicitly trained on.
A new generation of robotics.
General-purpose, not single-purpose. These are the commitments that shape how we build physical AI — and why a CALISTA robot is meant to be useful the moment it meets an unfamiliar task.
02
Full-stack by design
We build both the robots and the simulation-based training platform. Owning the whole stack is how perception, control, and learning stay in step.
03
Dexterity across tasks
From cracking an egg to seating a connector, the same hands handle delicate, contact-rich work with the precision each task demands.
04
Autonomy at normal speed
Tasks run autonomously and at a normal pace — not teleoperated, not slowed to a crawl. Useful speed is part of the definition of done.
05
Simulation that matches reality
Thousands of trials run in minutes, and what works in simulation works on the robot. Validation happens before deployment, not after.
06
Real-world intelligence, end to end
Our aim is to build real-world intelligence end to end — a new generation of robotics builders making highly capable, general-purpose machines.
Built for the real world
The same system, across very different rooms.
The tasks we demonstrate aren't confined to one setting. They span the home, the laboratory, and the assembly line — which is exactly what general-purpose is supposed to mean.
In the home
Everyday handling in an unstructured kitchen: cooking eggs, making a smoothie, picking up and putting away an assortment of objects.
Cook eggsSmoothieTidy & sort
In the laboratory
Precise, repeatable bench work: pipetting samples and handling instruments with the steadiness and accuracy that lab protocols require.
PipetteSample handlingProtocols
On the assembly line
Contact-rich manual assembly: routing and seating wire harnesses, picking multiple parts, and the fiddly steps that resist fixed automation.
Wire harnessMulti-pickConnectors
Inside the platform
How a new skill comes to life.
01Demonstrate
Capture the task
A new skill starts from a demonstration of the task and the environment it lives in — the goal, the objects, the constraints.
02Simulate
Thousands of trials in minutes
The platform runs the task across thousands of simulated variations in minutes, exploring grips, angles, and recoveries.
03Validate
Matched to real
Results are checked against real-world performance, so a skill is validated before it ever runs on hardware.
04Deploy
Out of the box
The skill joins the robot's shared library and runs autonomously, at normal speed — ready to generalize to the next task.
We're building real-world intelligence, end to end.
CALISTA is a new generation of robotics builders on a single mission: highly capable, general-purpose robots. We're backed by several venture capital firms and a number of prominent individual investors, and we're actively hiring across the technical roles that make physical AI real — perception, control, learning, simulation, and hardware.
Where we're pushing
The open problems we care about most on the way to general-purpose physical AI.
Generalization to the unfamiliar
Getting one system to handle tasks and environments it was never explicitly trained on — the difference between a demo and a general-purpose robot.
Dexterous, contact-rich manipulation
Closed-loop control that's gentle enough for an egg and exact enough for a connector, across the long tail of objects the world actually contains.
Sim-to-real fidelity
Closing the gap between thousands of fast simulated trials and what the hardware does, so validation in minutes predicts behavior on the floor.
Full-stack systems
Designing hardware, intelligence, and the training platform as one — because broad dexterity and reliability come from the whole stack moving together.
Let's build physical AI.
Whether you're exploring a deployment, the platform, or a role on the team — tell us what you're working on and we'll get back to you.
Call(900) 863-9266
Emailhello@calistapro.com
Visit2828 E Olympic BlvdLos Angeles, CA 90023
CompanyCALISTA LLCGeneral-purpose robotics