Skip to content
Physara
01 / Platform

A simulation stack built for the way robots actually move.

Physara starts with the physical details that matter: body movement, contact, balance, hands, sensors and actuators. Everything else is built around them.

Focus
Contact-rich dynamics
Compute
GPU-native parallel sim
Output
Policies + evidence
02 / Architecture

From a robot model to a testable skill.

The platform focuses on the parts that are hardest to reproduce on hardware: whole-body contact, dexterous hands, balance, collisions and recovery.

Physara simulation architecture
From a robot model to a testable skill
Fast parallel simulation
01 / Model
◒

Body model

Articulated structure, inertial properties, sensors and actuator behavior.

02 / World
⌁

Contact world

Surfaces, friction, collisions, terrain and randomized physical conditions.

03 / Learn
∿

Skill training

Train the same skill across many environments and conditions.

04 / Stress
↯

Stress testing

Test slips, pushes, balance loss and recovery before hardware trials.

05 / Transfer
↗

Best behaviour

Keep the best behaviours ready for controlled hardware testing.

Contact realisticHands detailedMotion modeled
03 / Training loop

Practise the difficult moments before hardware.

Create difficult situations, see how the robot responds, improve the skill, and test it again. The point is simple: more useful practice before physical testing.

How a skill improves
01

Build task

Scene + body
Ready
02

Change the conditions

Surface + objects
Active
03

Practise the skill

Repeated trials
2.48M eps
04

Test weak points

Slips + balance
Found 18
05

Check performance

New environments
93.7%
06

Prepare for hardware

Best behaviour
Ready
Progress / last 200k steps

Measure improvement where it matters.

The useful signal is not only success. It is whether the robot becomes more stable as conditions change.

SuccessInstability
0 stepsPolicy improves while risky contacts are surfaced200k steps
04 / Observability

Every trial tells you something.

A useful simulation should help a team understand what happened. Physara keeps the key events, movements and outcomes together so the next test is easier to plan.

Episode trace / recovery-004
01Foot contact0.00 s
02Disturbance injected0.42 s
03Center of mass shifted0.58 s
04Recovery torque0.74 s
05Stable stance1.06 s
Contact force 2.18 kNRecovery time 480 msOutcome Pass

Turn failed trials into better tests.

When a trial fails, teams can see the movement around the failure and decide what to change next: the skill, the environment or the robot model.

State joint + sensor tracesEvent contact + disturbance timelineOutcome success + stability metrics
05 / Transfer

Simulation supports the move to hardware.

The simulator gives a skill room to improve. The policy is tested in controlled conditions, then the best result can move toward a real robot.

01

Experience

Generate many physical variations around the same task, including failures that would be expensive to gather by hand.

Simulation
↓
02

Policy

Train, evaluate and compare candidate behaviors with the same telemetry and holdout scenarios.

Learning
↓
03

Hardware

Move the best-performing policy toward controlled physical trials and feed the evidence back into the next simulation cycle.

Validation
Design principleKnow what changes between simulation and reality.