Where to Slow Down: Risk-Aware Action Chunking for Robotic Margin Cutting

Surgeons already accept a trade: spending extra minutes on preparation and careful inspection buys cleaner tumor margins. Haptic boundary guides in breast surgery stretched resection from about 9.5 to 13.8 minutes but cut margin breaches. AI lesion labeling in endoscopy lengthened procedures from a median of 49 to 82 minutes, yet let novices match experts. Learned surgical robots spend time very differently. The leading imitation-learning policies use action chunking, meaning they predict and commit to a short block of future motions at once instead of one motion per camera frame. They commit to chunks of the same length everywhere along a task, whatever is at stake there. Meanwhile, computational pathology shows that what is at stake along a tumor margin is not uniform. Local recurrence tracks microscopic structure, such as fibrous bands that wall lymphocytes off from tumor nests and disordered cell neighborhoods. Those structures are patchy and spatially correlated, so they vary from one stretch of the margin to the next.

This note proposes a resection controller that decides, at each step along the cut, whether to keep committing to long, fast motion blocks or to stop and pay for local sensing. It makes that choice from the expected cost of a mistake at that location. That cost is estimated from a spatial risk map whose statistics come from real tumor cell-graph data and which is revealed only gradually by costly probes. In two sentences: the robot should move fast where an error is cheap and slow down and look where an error is expensive, and "expensive" should come from the spatial structure of the tumor's surroundings, not from the robot's own uncertainty. We test whether this beats uniform caution, uncertainty-triggered replanning, and simple distance rules on the trade-off between time and margin breaches, and whether the gain depends on the risk map being spatially structured.

Core Claim

Existing adaptive chunking methods decide how much of a predicted block to execute from the policy's own disagreement. Adaptive Action Chunking Strategy from World Feedback in Mixed Traffic samples 32 candidate chunks and truncates when they diverge. Mixture of Horizons in Action Chunking runs several horizon lengths in parallel and executes longer prefixes when the horizons agree. Both measure how unsure the policy is. Neither measures how costly an error would be at that location.

In margin cutting these two quantities come apart. A confident policy tracing a smooth contour can be heading into the one stretch of margin where a 1 mm deviation leaves a high-recurrence niche behind. Our claim is that the commitment decision should depend on consequence: the expected risk-weighted breach over the region a chunk would sweep, computed from a probabilistic estimate of the risk field. That estimate is useful only because tumor microenvironment risk is spatially correlated, so a few probes tell the robot about nearby tissue it has not yet sensed.

The decision rule can be stated compactly. At decision time $t$, the base policy proposes a chunk of length $k$ that sweeps tool positions $x_{t:t+k}$. The current posterior over the risk field is $q_t(r)$, and $p_{\text{dev}}(x)$ is the probability that the executed tool deviates past the planned margin at position $x$. We estimate $p_{\text{dev}}$ from the spread of sampled chunks, and it grows with $k$ because open-loop error accumulates. The controller chooses the action $a$ (execute $k$ steps for some $k$, or probe a local patch) that minimizes

$$ J(a) = \tau(a) + \lambda \, \mathbb{E}_{r \sim q_t} \Big[ \sum_{x \in \text{swept}(a)} p_{\text{dev}}(x)\, r(x) \Big], $$

where $\tau(a)$ is the wall-clock time the action costs and $\lambda$ trades minutes against risk-weighted breach. For a probe, the second term is the expected reduction in this breach cost on the next chunks after the posterior is updated. Sweeping $\lambda$ traces a time–quality frontier, which is our quantitative version of the clinical setup–execution trade-off.

Research Program

The base policy reuses an existing method. Surgical Robot Transformer (SRT) showed that action chunking with transformers, using a hybrid-relative action space (translation relative to the endoscope, rotation relative to the tool tip), reaches 90–100% success on dVRK tissue lifting, needle handover and knot tying. Absolute camera-frame actions failed completely there because cable slop causes up to 5 cm of end-effector error. We adopt that action space and train an ACT policy on teleoperated demonstrations of contour cutting around a lesion with a prescribed margin.

The risk side draws on two pathology papers, which motivate the problem and supply the map-building recipe. Cell graph analysis in hepatocellular carcinoma built minimum-spanning-tree cell graphs over about a million cells per slide and found that peritumoral fibrosis and community entropy predict one-year local recurrence (AUC 0.817). AI-powered spatial cell phenomics defined ten cellular niches by clustering neighborhoods within a 34 µm radius and found high-risk niche patterns in nearly half of Stage I patients. Neither paper connects these maps to any intraoperative action.

We will compute niche and entropy maps from whole-slide images, starting with TCGA-LIHC as used by Yuan et al. We then rescale them onto phantom geometry as risk fields $r(x)$. We are explicit that this is a proxy. A slide covers millimetres and the phantom covers centimetres, so what we transfer is the spatial statistics of real tumor-margin heterogeneity (patch sizes, correlation lengths, how high-risk regions cluster), not anatomical ground truth. The posterior $q_t$ is a Gaussian process whose kernel is fit to these maps. This is where spatial immune architecture enters the method instead of serving as decoration.

Probing models a real intraoperative cost, such as frozen section or a future virtual-histology readout like HEX, as a pause that reveals $r$ in a small patch. In simulation this is exact. On hardware we propose printing the risk field in ink visible only under UV light on gel-wax phantoms of the kind used by Connolly et al.. A probe switches on a localized UV view for a fixed time cost, so the information is real, physically registered and genuinely hidden until paid for.

Two other lines of work frame what we keep and what we change. Safety filtering follows SSP, which wraps policies in a control-barrier-function quadratic program with 3–10 ms latency around rigid no-go primitives. We keep a filter of that kind for hard constraints, but our contribution concerns graded, partially observed risk that a barrier function cannot express without knowing where the risk is. Front-loaded reconstruction follows From Monocular Vision to Autonomous Action, which spends 20–40 minutes reconstructing an airway tumor in 3D before a fixed polynomial cut (2.223 mm surface error). That work is the extreme "all setup up front, then blind execution" end of the trade-off. Our controller interleaves setup and execution instead.

Baselines

The naive reference is the SRT-style ACT policy with a fixed chunk length and no probing, cutting at the nominal margin. It gives the fastest time and an uninformed breach rate. Its uniformly cautious counterpart uses the shortest chunk length everywhere and probes on a fixed grid. That variant buys quality with time and anchors the slow end of the frontier.

The strongest existing competitor is uncertainty-gated chunk execution in the style of UA2C, transplanted onto the same ACT backbone: sample candidate chunks, truncate and replan when they diverge, and trigger probing when divergence is high. It is the closest published mechanism to ours and differs only in what drives the gate. The reported number for it is a 13.17% reward improvement over fixed-length chunking and an 18.37% reduction in cumulative jerk in mixed traffic. Mixture of Horizons, with 99% average success on LIBERO using $\pi_{0.5}$, is a second gated competitor; we implement its horizon-consensus head on our backbone. No reported result was found on Lacuna for chunked policies on margin cutting with heterogeneous risk, so all numbers in our domain will be ours.

We also include a geometric baseline in the spirit of Acar et al.: full up-front probing of the whole field, then a variable-offset planned contour followed by a classical controller. And we include the SSP-style barrier filter applied over the fixed-chunk policy with risk thresholded into hard zones.

Fairness rules apply to every method. All methods share the same demonstrations, the same base network and training budget, the same probe cost and the same hyperparameter search budget. Each method's free trade-off parameter is swept so that we compare full frontiers, not single points.

The key ablations remove each ingredient of $J$ in turn. One replaces the Gaussian-process posterior with a spatially uncorrelated prior, so probes inform only the patch they touch. Another drops the $p_{\text{dev}}$ dependence on $k$. We count the proposal as beating a baseline only if its frontier dominates: at matched total time, it must have lower risk-weighted breach with non-overlapping confidence intervals across held-out maps.

Why not just…?

Why not just set a spatially varying safety offset from the risk map and follow it with a planner? If the risk field were fully known before cutting, this would likely be enough, and the full-probing geometric baseline tests exactly that. The case for our controller rests on risk being only partly observable, with every probe costing time, which mirrors the clinical reality of 13.8 versus 9.5 minutes. The geometric baseline pays its entire sensing cost up front. If its frontier matches ours, the honest conclusion is that interleaving sensing with execution adds nothing, and we will report it that way.

Why not just use uncertainty-triggered replanning, which already exists? This is the most serious threat. In regions where the policy struggles, policy uncertainty and risk may be correlated, and then entropy gating would capture most of the gain. We test the alternative directly. We build held-out maps where high-risk patches fall on geometrically easy stretches of contour, where the policy is confident, and maps where they fall on hard stretches. Consequence gating should win only in the first case. If it does not win there, the core claim fails.

Why not just apply a hand-written rule, such as shortening chunks within a few millimetres of predicted high risk? Such a rule may recover much of the benefit, and we include it with its threshold tuned on the same budget. What it lacks is a principled decision about when to probe and how far probes generalize. If the rule matches our method, the contribution shrinks to the benchmark and the finding that risk-indexed commitment matters, which is still worth publishing.

Why not just ignore microscopic structure and use bulk density or a uniform margin? The pathology evidence argues that bulk counts miss what predicts recurrence: Yuan et al. report a 23.1% C-index gain over PatchGCN from graph structure. In our setting the shuffle test below directly measures whether spatial structure matters for control, not just for prognosis.

Decisive Checks

We will run the experiments first in a planar deformable-sheet cutting environment built alongside SurRoL. SurRoL's finding that realistic contact physics raised real-dVRK needle-pick success from 58% to 86% argues for modelling contact carefully rather than using distance thresholds. We then move to dVRK cutting on UV-ink gel-wax phantoms. Metrics are total task time, risk-weighted breach (the integral of $r$ over tissue left inside the intended margin), unweighted breach length, healthy tissue removed, safety-filter interventions and jerk.

The first decisive check is frontier dominance over the uncertainty-gated baseline on held-out pathology-derived maps, especially the "risk on easy geometry" split. The second is a shuffle control. We keep each map's marginal risk distribution but destroy its spatial correlation by permuting patches. If our advantage over grid probing comes from exploiting spatial structure, it should collapse on shuffled maps. If the advantage survives shuffling, then cell-graph architecture is not doing the work we claim, and we will say so.

The third check is transfer from simulation to the dVRK phantom: the frontier ordering among methods should be preserved. The fourth is clinically framed. At the operating point where our method matches the naive policy's time plus a budget comparable to the clinical overheads cited above, we report how much of the breach reduction achieved by full up-front probing it recovers.