CL-25 — mjlab continual learning

Extending the teacher → distillation continual-learning pipeline from 4-5 tasks to the full 25-task Class-A suite (Franka + 2-finger gripper).

Benchmark previews

Random-agent clips of every registered task — scene layout and asset placement, not task solvability.

37 clips

Phase 1 — classical teachers

Scripted-policy rollouts for the 25 Class-A tasks, each card showing the measured success rate against the 0.90 bar.

25 tasks in scope

CL-V2 — realistic assets

Every Class-A task rebuilt with a real, textured, real-scale object or mechanism (YCB, Google Scanned Objects, Poly Haven, or built): asset still + turntable, verified physics / init distribution / success predicate, and the classical teacher re-measured at n = 128 with the new asset.

25 tasks, all seven gates green

CL-V3 — teachers to ~100 %

The classical teachers re-strategised on the realistic assets, measured at n = 128 on a frozen initial-pose distribution (object yaw, mechanism yaw + height, robot joint noise); only tasks at or above 0.90 are published.

0 of 25 published