We present Video2SwimFish, an automated pipeline and benchmark for building controllable fish assets from real-fish videos for underwater embodied AI. Given synchronized multi-view videos of an individual fish, the pipeline generates a metrically scaled deformable mesh from a VLM-selected canonical frame, builds internal articulation adapted to that individual's morphology through a VLM actor-critic loop, and extracts a Biological Locomotion Manifold (BLM). The BLM provides a low-dimensional action space bounded by real-fish motion. Compared with learning directly in joint space, it lets the individual swimming policy of each reconstructed fish converge faster and reach a higher task success rate. We release two paired datasets: synchronized top- and front-view recordings of 120 individual fish across 6 species, and the controllable assets and individual swimming policies derived from them. Because every asset is tied to the animal it came from, the dataset supports a benchmark that evaluates locomotion learning not only on task success but on fidelity to that individual's locomotion behavior. We evaluate on two tasks, 3D trajectory following and target reaching, and a case study in which a simulated BlueROV underwater robot captures the assets. We further validate biological realism with a human study in which fish experts rate the simulated fish, as a complement to the automated metrics. We find that task success and locomotion fidelity do not necessarily improve together, and we identify faithful reproduction of individual animal locomotion as an open challenge for the community. Project website: https://hangongchen.github.io/video2swimfish-web/
翻译:暂无翻译