Robot policies stop reading only the current frame: optical-flow supervision predicts future motion, and a 0.87B model averages 63.3% on real dual-arm tasks
Lead
Evo-World adds optical-flow-supervised future motion prediction to a lightweight vision-language-action policy, reaching 91.1% tier-averaged success on Meta-World MT50, a 13.0% relative gain over the non-predictive baseline, and beating the evaluated baselines on AgileX single-arm and bimanual tasks with 0.87B parameters.
Story
A robot policy can now learn how the scene will move next, instead of only reading what the current frame shows. Earlier vision-language-action models trained on action labels alone, and action supervision did not explicitly require a policy's representations to anticipate scene change; existing future-observation prediction entangled task-relevant motion with largely static appearance. On Meta-World MT50, a variant predicting future RGB frames exceeds the non-predictive baseline by 3.7 percentage points, and the optical-flow-supervised Evo-World improves on that variant by a further 6.8 points under the same training and evaluation protocol.
The predictive signal enters action generation as a latent motion representation, so deployment needs no future image or optical-flow estimator. Encoders initialized from DINOv2 and CLIP encode the current view and instruction, a multilayer perceptron predicts instruction-conditioned demonstration motion, and a bounded FiLM combines the current and predicted representations to modulate the vision-language tokens. The training target is encoded from three-channel optical-flow renderings precomputed with RAFT and is used only during training; on an AgileX dual-arm robot each of floral arrangement, mug hanging, and microwave placement is evaluated for 20 trials.
What to watch
The report covers the simulation benchmarks Meta-World MT50 and RoboTwin, plus three real-world tasks on an AgileX dual-arm robot, with 20 trials per real-world task. It is aimed at researchers and engineers working on robot manipulation policies and lightweight deployment. The current motion targets describe 2D image-plane displacement without explicit 3D geometry, and future work will explore geometry-aware supervision.
The ablation conclusions are bounded by the same training budget and 100,000-step second-stage checkpoints on Meta-World. The linear-decoding analysis assesses end-effector displacement only and does not establish that the predictor adds linearly decodable motion information beyond the current representation. The motion targets come from 2D image-plane displacement without explicit 3D geometry.