CL-25 — mjlab continual learning
Latest: SI continual learning — one shared student across 15 tasks. Earlier galleries below retain their original recordings.
CL-V5 — SI continual learning
One shared student policy distilled sequentially across 15 certified RL teachers (>95% standalone success), using Synaptic Intelligence to resist forgetting. Two orderings, three seeds each, per-task videos.
CL-V4 — independent RL teachers
One PPO policy per task: a short progress summary, measured success, and a recorded RL-policy video for each task.
Benchmark previews
Random-agent clips of every registered task — scene layout and asset placement, not task solvability.
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.
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.
CL-V3 — corrected-physics teacher review
New videos with translucent task-shaped goals, corrected mechanism clearances, and arm/wrist/palm collisions. Updated teachers evaluated on 128 single episodes per task, including success and failure examples.