Every task. Already defined. Ready to train.
Each Kyros task ships with a reward function, success criteria, observation space, and evaluation harness. You write the policy. We handle the bookkeeping.
Manipulation
180+ tasksGrasping, pick-and-place, in-hand reorientation, tool use, deformable manipulation, multi-arm coordination.
Locomotion
80+ tasksBipedal and quadrupedal walking, stair traversal, recovery from disturbance, terrain adaptation.
Perception-Driven
120+ tasksVisual servoing, semantic grasping, dynamic obstacle avoidance, language-conditioned policies.
Assembly & Insertion
60+ tasksPeg-in-hole, screw fastening, cable routing, gear meshing — high-precision contact tasks.
Logistics
90+ tasksBin picking, packing, sorting, palletization across SKU diversity and clutter levels.
Whole-Body Control
40+ tasksCoordinated mobile manipulation, dynamic balance under load, reactive whole-body planning.
Each task is a complete spec.
No hand-rolled reward shaping. No undocumented success criteria. Every task is reproducible from the registry.
Reward function
Tunable, documented, version controlled. Default rewards have been validated against transfer outcomes.
Success criteria
Explicit pass/fail metrics. No ambiguity over what 'solved' means.
Observation & action spaces
Well-typed observation/action specs that match common robot platforms out of the box.
Evaluation harness
Run any policy against the standard evaluation suite. Get a transfer-likelihood score.