Yinpei/mme_vla_suite

This repository contains fine-tuned checkpoints for MME-VLA-Suite.

We train 14 VLA variants in total, all based on the $\pi_{0.5}$ model.

Name	Memory type	Memory representation	Memory integration	AVG Success	Released
symbolic-simple-subgoal	symbolic	simple subgoal	language concatenation	29.00	✅
symbolic-grounded-subgoal	symbolic	grounded subgoal	language concatenation	33.06	✅
perceptual-tokendrop-context	perceptual	token dropping	memory-as-context	34.50	✅
perceptual-tokendrop-modul	perceptual	token dropping	memory-as-modulation	38.04	✅
perceptual-tokendrop-expert	perceptual	token dropping	memory-as-expert	34.86	✅
perceptual-framesamp-context	perceptual	frame sampling	memory-as-context	30.68	✅
perceptual-framesamp-modul	perceptual	frame sampling	memory-as-modulation	44.51	✅
perceptual-framesamp-expert	perceptual	frame sampling	memory-as-expert	36.25	✅
recurrent-ttt-context	recurrent	TTT	memory-as-context	22.28	✅
recurrent-ttt-modul	recurrent	TTT	memory-as-modulation	21.97	❌
recurrent-ttt-expert	recurrent	TTT	memory-as-expert	22.35	✅
recurrent-rmt-context	recurrent	RMT	memory-as-context	19.46	❌
recurrent-rmt-modul	recurrent	RMT	memory-as-modulation	20.17	❌
recurrent-rmt-expert	recurrent	RMT	memory-as-expert	18.15	❌

We release all symbolic and perceptual memory MME-VLA variants for research.
Due to recurrent memory currently underperforms, we release only a subset. We will release newer recurrent variants once we obtain better results.