A robot world-action model shrinks to 0.88B: 95.5% LIBERO success at 57 ms

A pretrained video world-action model can be made much smaller while keeping most of its control success and still predicting future frames. Earlier compact policies either predict only actions and drop the world model, or skip future prediction at test time; the teacher Cosmos Policy reports 98.5% success on LIBERO in the introduction and 98.2% in the table and Q1, at 624 ms, 42.6 TFLOP and 7.55 GiB per 16-action chunk. Across four LIBERO suites and 10 trials per task, 400 episodes, SmolWAM succeeds in 382 (95.5%, 95% Wilson interval 93.0-97.1) against the teacher's 393 (98.2%, 96.4-99.1), both from the same initial states; no difference is detected on Spatial and Object, and the deficit lies in the multi-step Goal (94 vs. 99) and Long (89 vs. 97) suites.

Most of the saving comes from replacing an inference interface built for video generation with one built for control, not from discarding the model weights. The teacher used a 4.8B-parameter T5-XXL text encoder, a VAE that encoded a 33-frame clip per query, and five denoising steps of a 2B-parameter transformer over 1,764 tokens. Running the teacher's unmodified weights for one step on single-frame inputs cuts latency 5 times and compute 5.7 times, to 126 ms and 7.5 TFLOP, with no loss on the screening set; SmolWAM's trunk redesign adds a further 2.2 times in latency, 3.3 times in compute and 1.9 times in memory, and of its 57 ms the transformer takes 37, the VAE encode 13 and preprocessing 7.

The compressed world head still forecasts motion rather than texture. It was unclear whether future prediction survives when a video world model is compressed. On 96 held-out contexts, SmolWAM's forecast of the view 16 steps ahead lowers latent L1 by 24% relative to copying the current frame (0.101 vs. 0.133) and lowers the end-point error between RAFT optical flow to the predicted and to the real future from 1.63 to 1.43 pixels; decoded forecasts keep the scene and place the arm near its future pose but blur its appearance, and LPIPS stays at the copy-frame level (0.128 vs. 0.115).

SmolWAM is evaluated on LIBERO with the teacher's observation pipeline: wrist and third-person images at 224 px, proprioception, and 16-action chunks executed open loop, seed 195, 10 trials per task for 400 episodes; latency is wall-clock time on one A100-SXM4-40GB at batch 1 and bf16, including preprocessing and VAE encoding, with text features cached for every model. The authors propose extending evaluation to additional robotics benchmarks such as RoboCasa and, ultimately, to real-robot settings.

The authors state that evaluation is limited to LIBERO, and that reducing transformer depth disproportionately affects long-horizon tasks, with the largest degradation on LIBERO-Long, so pruning strategies that preserve long-horizon control remain an open problem. Validation action L1 settles near 0.09, ten times the teacher's own one-step error, while closed-loop success stays at 95.5%, so offline L1 is a poor proxy for closed-loop success; 7,500 further training steps lower the training loss but neither the validation loss nor success. Porting ABot-M0.5 components yields SmolWAM-LA at 53.3% screening success and 239 ms, and reward fine-tuning of its world head collapses success to 5.0-16.7% within 300 steps while worsening every image metric. In single-flag ablations every change lands within five episodes of the 78.3% reference, so at that budget the screen cannot separate these choices, and six planned ablations were not run.