BEHAVIOR-1K research, facts checked Oct 29 HST
A world action model (WAM) starts from a video generation model instead of a vision-language model and learns to predict future video and robot actions together. This page compares six of them for one question. Could one replace pi0.5 as our BEHAVIOR-1K policy next year? Flex-pi is the strongest fit today, and the Nov 4 challenge results should come first.
Every fact was checked against the paper full text, the official repo or the model card on Oct 29 HST. Anything marked UNVERIFIED was not found in a source. The target is a Galaxea R1Pro (two arms, mobile base, torso) with three cameras (head plus two wrists), a 23-dim action, tasks several minutes long, and a 24 GB GPU at test time.
| model | size and backbone | imagines video at test time | speed (GPU) | fits 24 GB at test | code and weights license | fine-tune cost | vs pi0.5 (same benchmark) | mobile base |
|---|---|---|---|---|---|---|---|---|
| Flex-pi | 6B, Wan2.2-5B | optional per call | 60 ms action-only, 193 ms with video (RTX 5090) | yes, 16 to 26 GB | MIT, pretrained weights released | LoRA fits one 24 GB GPU (16 to 19 GiB peak), full needs 4 to 8 x 80 GB | RoboTwin 94.6 vs 79.8, LIBERO-Plus 88.6 vs 85.7, real bimanual 3.5x | no |
| Fast-WAM | 6B, Wan2.2-5B plus 1B action expert | no | 110 to 190 ms (paper and README disagree on GPU) | likely, VRAM UNVERIFIED | code MIT, weights LIBERO and RoboTwin only, license UNVERIFIED | 8 GPUs 20k steps (LIBERO), 64 GPUs 30k steps (RoboTwin) | RoboTwin 91.8 vs 79.8, LIBERO 97.6 vs 96.9, lost to pi0.5 on real towel folding (Galaxea R1 Lite) | no |
| LingBot-VA | 5.3B, Wan2.2-5B plus 350M action stream | yes, every chunk | not reported | borderline, about 24 GB with VAE and text encoder on CPU | Apache 2.0, pretrained base released | about 3K steps on 8 GPUs post-training, LoRA in LeRobot | RoboTwin 92.9 / 91.6 vs 82.7 / 76.8 | no |
| Cosmos Policy | 2B, Cosmos-Predict2-2B | yes, jointly with actions and values | 0.16 s at 1 step, 0.95 s at 10 steps (H100) | yes, 6 to 10 GB | code Apache 2.0, weights NVIDIA noncommercial | 48 h on 8 to 64 H100, no LoRA path | LIBERO 98.5 vs 96.9, ALOHA real 93.6 vs 88.6 (but 89.3 vs 92.5 out of distribution) | no |
| GlanceWAM | 1.6B, SkyReels-V2-1.3B | one frame 3 s ahead, in the background | 48 ms (A100 40 GB) | likely, VRAM UNVERIFIED | code MIT, weights inherit SkyReels terms (UNVERIFIED) | 4 H200 for 10K to 15K steps | LIBERO 99.0 vs 96.9, real robot 79.2 vs 69.6 | no |
| DreamZero | 14B, Wan2.1-I2V-14B | yes, jointly with actions | 150 ms only on 2 GB200 fully optimized, about 3 s on H100 | no, weights alone about 45 GB | code Apache 2.0, DROID weights noncommercial | pretraining compute UNVERIFIED, 50K-step task fine-tunes, paper says LoRA is worse | 62.2% task progress vs 27.4% for the best VLA (which VLA is UNVERIFIED) | yes, AgiBot G1 |
Flex-pi. The strongest fit today. It is the only one with a documented path to both run and fine-tune on a 24 GB GPU, its license is MIT for code and weights, and pretrained checkpoints are public. Its action-only mode is faster than pi0.5 and its video mode can be switched on to inspect what the model expects. Its gains over pi0.5 are large on RoboTwin and real bimanual tasks. Its rig is stationary, so the R1Pro base and torso channels would be new outputs that only our demos teach.
Fast-WAM. Close behind on paper, with the same 6B Wan2.2 size and no test-time imagination. It has no embodied pretrained checkpoint, an unstated weights license, and it lost to pi0.5 on the only Galaxea robot task it reports.
LingBot-VA. The only one someone has already prepared for BEHAVIOR. The dataset AkashKarnatak/lingbot-va-b1k holds Wan2.2 latents of the head and both wrist cameras for all 100 tasks (5,000 train and 1,000 val episodes from the 2026 demos). Its model repo is private, so whether it was submitted or how it scored is unknown. It imagines video every chunk, which costs time and fills a 24 GB card.
Cosmos Policy. Small and fits the GPU easily, but fine-tuning used 8 to 64 H100 for two days with no LoRA path, the weights are noncommercial, and it was worse than pi0.5 on out-of-distribution ALOHA.
GlanceWAM. The cheapest and fastest. It trains on 4 H200, which KOA has, and it beats pi0.5 on LIBERO and a real robot. It has no mobile or long-horizon results, and the batch size differs between the paper and the README.
DreamZero. The closest match to the R1Pro, since the AgiBot G1 is a mobile bimanual robot with three cameras, a 22-dim action including base velocity, and about 5-minute episodes. It cannot run on a 24 GB GPU at 14B, and its 5B variant scored 21% against 50% task progress.
If we try a WAM, start with Flex-pi. A LoRA fine-tune from its pretrained checkpoint fits one KOA GPU, inference fits the 24 GB limit, and the license is clean. GlanceWAM is the low-cost second option on 4 H200.
Wait for the Nov 4 results and any WAM team reports first, since they will show whether WAMs beat pi0.5 on BEHAVIOR at all. A JEPA-WAM style auxiliary loss on pi0.5 stays the cheaper bet that keeps our existing checkpoints.