Source-linked AI summary
BladeMaster: Real-Time Robotic Cutting Simulation with Online-Generated Persistent Discontinuities
Zhanyu Yang, Yunuo Chen, Yanjia Huang, Joseph Masterjohn, Yin Yang, Chenfanfu Jiang
TL;DR
Robotic cutting simulation must generate cuts as the blade moves and preserve separation when cut faces later interact. BLADEMASTER uses persistent particle side labels in a GPU-accelerated TLMPM framework to represent those cuts, and experiments report cutting followed by manipulation at faster-than-real-time rates on representative tasks.
Problem
Robotic cutting simulators must represent tool-driven cuts and preserve discontinuities after withdrawal, while prior approaches can require predefined surfaces or tie separation to tool geometry.
Source: §I · p. 1Method
BLADEMASTER uses swept-blade-generated persistent particle side labels to govern TLMPM grid coupling, with contact and two-way tool-material coupling for post-cut interactions.
Source: §I · pp. 1–2Results
Simulations demonstrate cutting and subsequent manipulation without artificial grid coupling across completed cuts, with faster-than-real-time performance on representative tasks and qualitatively similar real-robot separation and manipulation.
Source: §V · pp. 7–8Takeaways & Limitations
BLADEMASTER is presented as a step toward efficient policy learning and motion optimization for robotic cutting tasks.
Source: §V · p. 8The model represents topological separation only along blade-induced cuts and does not capture stress-driven crack initiation, propagation, or failure under non-cutting loads.
Source: §V · p. 8
Abstract
from arXiv · showhide
Cutting changes both the shape and topology of deformable objects, making accurate simulation challenging for robotic manipulation. A simulator must track the cutting tool as a cut develops, preserve the resulting discontinuities after tool withdrawal, and enable newly exposed surfaces to interact with the tool and with each other. Existing formulations often prescribe cut surfaces in advance or couple material separation to auxiliary geometric fields. We introduce BladeMaster, a GPU-accelerated cutting framework based on the total Lagrangian material point method (TLMPM). Our key idea is to encode the cutting history directly on material points through persistent side labels generated online from the blade geometry. These labels govern particle-grid coupling, preserving connectivity within intact material while preventing spurious coupling across cut faces after tool withdrawal. Our formulation supports progressive and intersecting cuts without predefined cut surfaces or particle duplication. Material-material contact enables cut surfaces to recontact and slide against each other without reconnecting, while two-way tool-material coupling allows material reaction forces to influence tool motion. Experiments demonstrate tool-driven cutting followed by manipulation, with faster-than-real-time performance on representative tasks.
I. INTRODUCTION
Robotic cutting simulation must generate cuts as a tool moves and preserve material separation during later contact and manipulation. BLADEMASTER uses persistent particle labels to represent that evolving separation, with contact and tool coupling for subsequent interaction.
II. RELATED WORK
- DiSECt requires a predefined cut surface, while CRESSim-MPM does not encode cumulative connectivity changes from successive cuts in the material state.Other MPM approaches also leave limitations after tool withdrawal or do not distinguish fragment identities in reported transfers.Source: §I · p. 1, §II · p. 2
I. INTRODUCTION
- BLADEMASTER assigns persistent side labels online from swept blade geometry to govern grid coupling and preserve separation after tool withdrawal.It supports progressive and intersecting cuts without predefined cut surfaces or particle duplication.Source: §I · pp. 1–2
- Material contact lets cut surfaces recontact and slide without reconnecting, while two-way tool coupling transfers material reaction forces to the tool.This supports manipulation of separated pieces after cutting.Source: §I · p. 2
- The framework is evaluated in cutting and manipulation scenarios, including comparison with prior methods, physical-robot behavior, and measured cutting resistance.The GPU implementation runs faster than real time on representative tasks.Source: §I · p. 2
III. METHOD
The method couples a deformable TLMPM body to a moving rigid cutting tool and augments particle state with cutting-progress and side codes. Swept blade geometry updates those codes to control grid coupling and retain cut separation.
III. METHOD
- The deformable body is discretized into material points and advanced with TLMPM, while a rigid-body system handles tool actuation, constraints, and state updates.The contact solver applies impulses to particles and returns the corresponding reaction wrench to the tool.Source: A. System and Notation · p. 2
- Particles store a cutting-progress variable and cut-side code in addition to their standard TLMPM state.Cutting progress resets at the start of each new cut, and contact impulses are scattered to the grid each substep.Source: A. System and Notation · p. 2
- Valid tool contact advances cutting progress, while the swept blade band assigns side codes that govern particle-grid sharing from the next substep.The blade sweep also disconnects intersected reference-grid links and triggers local surface reconstruction.Source: A. System and Notation · p. 3, Figure 2 · p. 3
B. TLMPM Discretization
The discretization uses fixed reference-grid neighborhoods so changes in material connectivity can be controlled through cut-side codes. Cutting resistance is updated from directional contact work, while swept blade geometry assigns persistent side labels.
B. TLMPM Discretization
- TLMPM evaluates interpolation weights and gradients at fixed reference positions, keeping particle-grid stencils constant and precomputable.The formulation uses an explicit MUSL scheme; compatibility is enforced during particle-to-grid transfer and velocity remapping.Source: B. TLMPM Discretization · p. 3
- Cut-side codes determine the logical grid receiving each particle’s mass, momentum, and internal-force impulse, controlling connectivity within fixed neighborhoods.The constitutive model is independent of the cutting formulation.Source: B. TLMPM Discretization · p. 3
C. Cutting Resistance and Side Assignment
- Particle position relative to blade-plane normal, finite edge axis, and front-back direction defines local blade coordinates for the cut.These coordinates support defining the finite cutting band.Source: C. Cutting Resistance and Side… · p. 3
- The method accumulates directional blade-contact work within the cutting band and releases the opposing impulse component when the empirical threshold sets cutting progress to 1.Cutting progress affects this contact component, not constitutive stress.Source: C. Cutting Resistance and Side… · p. 3
- An unassigned particle behind the blade edge receives a side label, which changes grid compatibility and contact selection beginning with the next TLMPM substep.Label assignment does not immediately change particle position, velocity, or stress.Source: C. Cutting Resistance and Side… · pp. 3–4
D. Topology-Aware Grid Transfer
Persistent per-cut side labels encode each particle’s cutting history and determine which logical-grid states can couple. Compatible material remains connected, while conflicting cut sides evolve independently and can still interact through explicit contact.
D. Topology-Aware Grid Transfer
- Each particle stores a ternary digit per cut, and later cuts add digits without overwriting earlier labels, supporting progressive and intersecting cuts without particle duplication.A digit changes only from 0 to 1 or 2; existing nonzero digits persist.Source: D. Topology-Aware Grid Transfer · p. 4
- Codes conflict when they assign opposite sides of any cut; an unassigned 0 remains compatible with either side, keeping intact material coupled ahead of the blade.Behind the blade, opposing cut faces remain separated.Source: D. Topology-Aware Grid Transfer · p. 4
- Compatible nodal states exchange momentum through antisymmetric pairwise exchanges, which conserve momentum and vanish when the states already share a velocity.Source: D. Topology-Aware Grid Transfer · p. 4
- Particles scatter to and read from their own logical grids, while compatibility coupling is applied during both initial P2G transfer and MUSL re-scattering.Conflicting grids remain independent even with overlapping reference stencils; explicit contact permits cut faces to recontact, and the cutting-band width is hk = 6∆x.Source: D. Topology-Aware Grid Transfer · p. 4
E. Contact Model
The contact model handles material, cut-face, and tool interactions with a shared impulse-based formulation. Contact is detected in the current configuration, then impulses enter the fixed-reference TLMPM update.
E. Contact Model
- A unified Coulomb-friction and impulse-scattering procedure handles inter-object contact, self-contact, cut-face recontact, and rigid-tool contact.Contact is detected in the current configuration, while impulses are scattered through fixed reference stencils into the TLMPM update.Source: E. Contact Model · pp. 4–5
- For self-contact, nearby reference neighbors are excluded because continuum coupling already connects them, but conflicting cut-side pairs remain candidates for crack-face contact.The contact pair is represented using particle spheres and predicted overlap.Source: E. Contact Model · p. 4
- Persistent contacts store tangential spring history for static friction, switching to kinetic friction when the static Coulomb limit is exceeded.The same history-based friction model is used for material–material and material–rigid contact.Source: E. Contact Model · p. 4
- Tool contact scatters the final impulse to the particle’s logical grid and returns equal-and-opposite force and torque to the tool.That same impulse updates material motion and supplies the contact work used by the cutting-resistance model.Source: E. Contact Model · p. 5
F. Reference-Grid Surface Reconstruction
The surface reconstruction uses a reference-grid link representation to create visible cut faces, then embeds those surfaces in side-aware TLMPM motion fields. This separates visual topology updates from mechanical surface motion.
F. Reference-Grid Surface Reconstruction
- Visible boundaries are reconstructed on a uniform reference grid using a linked-volume representation and local dual contouring.Source: F. Reference-Grid Surface Reconstruction · p. 5
- Blade-swept links are disconnected, and connected components of the remaining links define local material regions whose surface representatives form coincident faces across a cut.Intersections with original boundaries or blade-generated cuts are stored as Hermite points and normals; only affected cells are updated.Source: F. Reference-Grid Surface Reconstruction · p. 5
- Each surface representative inherits its region’s cut-side code and follows the corresponding compatible TLMPM motion field through fixed trilinear weights.Link connectivity determines visual topology, while side-aware embedding keeps surface motion consistent with the mechanical discontinuity.Source: F. Reference-Grid Surface Reconstruction · p. 5
G. GPU Implementation and Rigid-Tool Interface
The GPU implementation parallelizes the simulation kernels and scales grid storage with active topology states. Its rigid-tool interface can exchange impulses across different update rates.
G. GPU Implementation and Rigid-Tool Interface
- Stress evaluation, TLMPM transfers, contact handling, logical-grid aggregation, and cutting-state updates are implemented as GPU kernels.Source: G. GPU Implementation and Rigid-Tool… · p. 5
- Active topology codes are compacted into contiguous physical grid slots, so memory and computation scale with active states rather than the full 3K code space.Source: G. GPU Implementation and Rigid-Tool… · p. 5
- When the rigid-body system updates the tool less frequently than TLMPM, pose and twist stay fixed between updates while reaction impulses accumulate into an interval-averaged wrench.The returned wrench preserves net linear and angular impulses.Source: G. GPU Implementation and Rigid-Tool… · p. 5
IV. EXPERIMENTS
The experiments test online cut generation with a freely moving knife and whether cuts remain mechanically distinct after tool withdrawal.
A. Comparison with Existing Cutting Mechanisms
- A 0.1 kg knife initially tilted 25° falls freely; material reactions deflect its path, while lower knife sharpness leaves the block uncut.The online cut follows the evolving trajectory rather than a predefined surface.Source: A. Comparison with Existing Cutting… · p. 5, Figure 3 · p. 5
- After two radial cuts and knife withdrawal, BLADEMASTER lifts only the slice, while the single-field baseline also lifts surrounding material.The comparison uses identical particle sampling and knife motion; persistent side labels preserve mechanical separation after withdrawal.Source: A. Comparison with Existing Cutting… · pp. 5–6, Figure 4 · p. 6
- 0.654 mm vs 4.059 mm mean gap width halfway through the first cut, and 0.506% vs 8.746% middle-piece volume contraction after both cuts, favor BLADEMASTER over CRESSim-MPM.After withdrawal, only the left piece moves away in BLADEMASTER; shared nodal velocities move the other pieces in CRESSim-MPM.Source: A. Comparison with Existing Cutting… · p. 6
B. Cutting Scenarios
The scenarios demonstrate multiple cutting and manipulation behaviors, including physical comparisons of banana-cutting resistance and piece motion.
B. Cutting Scenarios
- Figure 5 pairs the first-cut, completed-cut, and post-push states for BLADEMASTER and CRESSim-MPM; only the left piece receives an initial leftward velocity.The final states are shown after the same elapsed time, with colored bands marking cut sides and ground markers providing position references.Source: Figure 5 · p. 6
- Three freely moving knives launched at 0.5, 0.8, and 1.1 m s−1 are decelerated by material reactions, with higher initial speeds producing greater penetration depths.Their trajectories are not prescribed, and the three cuts are simulated concurrently.Source: B. Cutting Scenarios · p. 6, Figure 6 · p. 6
- With the same cross-shaped in-plane motion, a shallow pass makes light surface scores while a deeper pass produces more pronounced intersecting partial cuts.Both cut paths intersect at the center without passing through the body.Source: B. Cutting Scenarios · p. 6, Figure 7 · p. 7
- Successive cuts release slices that topple onto the board and previously cut slices, forming an overlapping pile while uncut material remains connected below the blade.The sequence demonstrates progressive cutting and contact among the resulting pieces within one task.Source: B. Cutting Scenarios · p. 6, Figure 8 · p. 7
C. Real-world experiment & Cutting Resistance Evaluation
- Before knife–support contact, simulated and measured banana-cutting resistance forces are comparable in magnitude; averaging 20 samplings and refining resolution smooth the traces and improve their match.The resistance estimate is compared with a real robot’s joint-torque measurements during a single vertical banana cut.Source: C. Real-world experiment & Cutting… · pp. 6–7
- In a real banana-cutting experiment, simulation closely matches the physical experiment in the separation and displacement of pieces across the illustrated stages.The same end-effector trajectory is used in simulation and the physical experiment, with the knife pushing pieces aside after each cut.Source: C. Real-world experiment & Cutting… · p. 7
D. Runtime and Scalability
Representative cutting examples run faster than real time; the conclusion also reports qualitatively similar cut-piece separation and manipulation in a real-robot experiment.
D. Runtime and Scalability
- Fig. 9 plots banana-cutting axial resistance for a single run, a 20-run pointwise mean, and a refined run; late peaks mark knife contact with the rigid support.Total resistance includes material and rigid-support reactions.Source: Figure 9 · p. 7
- All listed cutting examples run faster than real time, with reported timings covering the simulation loop but excluding surface reconstruction and rendering.Table I defines wall/sim. values below 1 as faster-than-real-time execution.Source: D. Runtime and Scalability · p. 7, Table I · p. 7
- At a 2 ms time step, Fig. 10 compares runtime across particle and cut counts and reports occupied compact-grid slots, which are fewer than the full 3K code space.The figure caption specifies four particle counts for the runtime ratio and compact-grid slot counts at each cut count.Source: D. Runtime and Scalability · p. 7, Figure 10 · p. 7
V. CONCLUSION
- A real-robot experiment shows qualitatively similar separation and manipulation of the cut pieces.Source: §V · p. 8
- The method does not model stress-driven cracking or failure under non-cutting loads, and its spherical-proxy tool contact does not strictly guarantee nonpenetration.Active compact-grid states may still grow with the number and arrangement of cuts.Source: §V · p. 8
APPENDIX MOMENTUM CONSERVATION
The appendix establishes momentum preservation for the coupling in Eq. (8), under stated assumptions on particle masses, cut-side codes, and transfer weights.
APPENDIX MOMENTUM CONSERVATION
- The coupling in Eq. (8) preserves momentum in PIC/FLIP transfer and MUSL re-scatter of for any common blend ratio β ∈ [0, 1], assuming fixed particle masses mp and cut-side codes cp within a substep.Source: APPENDIX MOMENTUM CONSERVATION · p. 8
- Pairwise momentum exchanges cancel between compatible grid pairs, yielding momentum preservation at each occupied node and across the coupled grids.Source: APPENDIX MOMENTUM CONSERVATION · p. 8
- The coupling also preserves a common velocity shared by compatible occupied grids at a node because their pairwise momentum exchanges vanish.Source: APPENDIX MOMENTUM CONSERVATION · p. 8