Surgical tissue forecasts anchor to the last observation: median error ratio 0.912 across 12 clips, interval includes parity
Synopsis
Surgical robots need to anticipate how instrument motion will deform tissue when visibility is interrupted, and prior differentiable physics simulation predicted absolute tissue positions, so reconstruction error, initialization, and simulation offsets left the whole surface persistently displaced. The forecast now anchors each tissue point to its last valid observation and predicts only the deformation induced by the known future tool trajectory, using mesh-free position-based dynamics with scene-specific rigidity, attachment, and grasp coupling. Across 12 surgical manipulation clips, the referenced forecast has a median within-clip error ratio of 0.912 against the persistence baseline, 95% interval 0.787–1.070, improving 8 of 12 clips, with the interval including parity.
Interpretation
Tissue deformation forecasting now predicts only the displacement relative to the last observation instead of absolute tissue positions. Differentiable simulation previously output absolute positions, so reconstruction error, initialization error, and sim-to-real mismatch left the predicted surface spatially offset even when its deformation trend was informative. Across 12 surgical manipulation clips, the referenced forecast has a median within-clip error ratio of 0.912 against the persistence baseline, 95% interval 0.787–1.070, improving 8 of 12 clips, with the interval including parity; against tool-driven ridge the median ratio is 0.955, interval 0.911–1.048.
The physical model uses mesh-free position-based dynamics and identifies a small set of effective coupling parameters per scene to turn instrument motion into tissue displacement. Position-based dynamics was previously used mainly for stable updates with prescribed constraints and stiffness, rather than parameters identified from observation. Each scene uses a fixed reference point cloud, 12-neighbor elastic patches, and regional scaffold nodes, fitting regional rigidity, region scales, per-point attachment, and shared grasp adherence; the tool input is tracked jaw translation, with the grasp set and each grasp point's initial offset fixed throughout the episode.
An optional residual model captures the periodic motion left after physics fitting, using graph-spectral spatial modes with harmonic temporal dynamics. Periodic physiological motion independent of tool input was previously unmodeled and remained in the fitting error. The residual is also referenced to the last observation; on three manipulation pilots, conditional residual selection is modestly better than physics alone in 2 of 3 pilots, while joint fitting increases error relative to its physics component on all 3.
Perspective
The evaluation covers 12 development-selected surgical clips with usable tracked surfaces and grasp motion, from 10 surgeries, two of them calibrated stereo clips that support physical-unit errors. Forecasts are scored against offline reconstructed tracks, and fitting and scoring share the same depth, tracking, smoothing, and visibility selection. Periodic residual studies cover only three manipulation pilots and one passive recording.
Clip-level results vary widely, with the two calibrated stereo clips showing the largest improvements, and one of them limited to a 0.72 s forecast horizon. Residual gains on manipulation clips are clip-dependent and more limited under nested selection. In passive periodic forecasting, extending the observation window from 3 s to 37 s reduces the median harmonic-to-persistence error ratio from 5.540 to 0.489, but errors vary nonmonotonically with window length and one of three cutoffs worsens at 37 s. The three manipulation prefixes contain only 0.50, 1.19, and 2.47 fitted cycles, all frequency estimates lie at the search-band boundaries, and no independent respiratory signal is available to relate these oscillations to respiration.