Robot policies are being measured by compact foresight and deployable control

ConsisVLA-4D treats spatial consistency as an inference budget problem. It keeps 32 instruction-relevant visual tokens, aligns them across multiple camera views, and stores geometry in compact latent tokens. The reported gains are tied to both accuracy and speed: 21.6% better performance and 2.3× faster inference than OpenVLA on LIBERO, plus 41.5% better performance and 2.4× faster inference on real robot platforms.

T³VF addresses a different failure point in visual-foresight VLA models. It compares a predicted future image with the later observed image, then updates only learnable query tokens when action variance is low. On LIBERO-Plus with perturbed training, it raises Mantis average success from 49.3% to 52.1%, with larger gains on camera and lighting perturbations.

The latent-action study gives a controlled comparison that many VLA papers have lacked. Using one Qwen3-VL-2B-based baseline, it tests image-based latent actions and action-based latent actions across four integration methods. The strongest results depend on task type. LA-Direct reaches 96.6% on LIBERO-Long, a 10.8-point gain over the baseline. LA-Tok reaches 78.0% average success on RoboTwin 2.0, a 17.5-point gain, and improves motor-heavy tasks such as Move Can Pot from 46% to 70%.

The practical point is clear: latent actions are useful when they are matched to the supervision problem. Image-derived tokens help long-horizon scene reasoning. Action-derived tokens help normalize motor control across heterogeneous robot data.

Dream-MPC applies the deployment theme to continuous control. It samples only five candidate action sequences from a policy prior, rolls them through a learned latent world model, and takes one gradient step using predicted reward and terminal value. An uncertainty penalty discourages plans that enter model-error regions, and action reuse carries optimization work across model-predictive-control steps.

The efficiency claim is concrete. In the reported setup, Dream-MPC uses 15 world-model evaluations per time step, compared with 9,216 for the cited MPPI configuration. With BMPC across 24 continuous-control tasks, it improves IQM normalized score by 26.7% and mean normalized score by 20.5%. With TD-MPC2, it improves over the policy-only baseline but does not consistently match TD-MPC2 with MPPI.