The training sequence matters as much as the world.
A robot trained on the wrong progression learns shortcuts that fail on hardware. Kyros curricula package staged task difficulty, randomized conditions, and transfer checks into one reproducible training path.
curriculum_graphstage gates + eval suites
research benchbenchmark paper
deltaablationtransfer
physics
policy
transfer
5
progressive stages from clean task to transfer stress test
20+
pre-built curricula across manipulation, locomotion, and perception
Fork
adapt any sequence to your robot, task, or lab constraints
Anatomy of a curriculum.
A sequential 5-stage progression ensuring robust sim-to-real transfer.
90%
Stage 01: Controlled Primitive
Success Threshold
Robot encounters one object in a fixed position. Zero variation to build core primitives.
success threshold: 85%
Stage 02: Multi-Object Fixed
Three objects, fixed layouts. Robot learns to isolate target coordinates and ignore adjacent clutter distractors.
success threshold: 82%
Stage 03: Random Position
Single object, fully randomized coordinate offsets. Core spatial generalization starts across workspace.
success threshold: 78%
Stage 04: Clutter & Noise
Three objects in fully random locations with distractor blocks. Model isolates correct objects and plans collision-free paths.
success threshold: 70%
Stage 05: Domain Randomization
Lighting, friction, mass, materials, camera positions — all randomized. Surviving policy generalizes immediately to real hardware.
Light & Shadows
L
F
Friction CoeffMass & Inertia
M
Train in the right order.
Get early access to the full curriculum library.