PufferLib 5.0: Five Coolest RL Results, from Multitask Drone Policies to Selfplay Robocode Champions
jsuarez · x · 2026-09-14
jsuarez rounds up the five most impressive RL results in PufferLib 5.0:
- Multitask drone policies: a single policy can both race and arrange into multiple formations, built by an intern.
- Robocode champions via selfplay: the team ported the classic tank-coding game and its strongest historical bots to a high-performance C PufferEnv. The standout result: agents trained purely via selfplay, never facing the reference bots directly, still beat every bot they train against.
- Craftax SOTA (post truncated).
The writeup highlights PufferLib's selfplay generalization in high skill-ceiling environments and multitask policy training.
Related event: PufferLib 5.0 Sets New RL Records on Craftax and NetHack(4 posts)→
More from Research
- DeepMind's Economics of Intelligence: AI reshapes economy via agent coordination, not solo model smarts — Scobleizer · 2026-10-31
- Nesformer: an entire Nintendo crammed into ~1GB of Qwen3.5 weights, following the Qwenputer method — idanbeck · 2026-10-31
- SCOPE workshop proposed for ICLR 2027 asks what counts as an AI scientific discovery — mariyaivasileva · 2026-10-31
- New MolmoSpaces-CommonSense benchmark: no robot policy tops 30% on manipulation tasks — RanjayKrishna · 2026-10-31
- Mischief launches video model trained from lab-grown brain cells at 0.5¢ per second — evilsocket · 2026-10-31
- Stanford CS336 notes: Percy Liang on scaling recipes and hyperparameter transfer — stanfordnlp · 2026-10-31