INF-3DP: Implicit Neural Fields for Collision-Free Multi-Axis 3D Printing

Aoran Lyu; Chengkai Dai; Dezhao Guo; Guoxin Fang; Hailin Sun; Jiasheng Qu; Yeung Yam; Zhuo Huang

REVIEW 4 major objections 5 minor 4 references

INF-3DP: Implicit Neural Fields for Collision-Free Multi-Axis 3D Printing

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read INF-3DP claims that replacing every discretized stage of multi-axis 3D printing with implicit neural fields yields a single differentiable pipeline for toolpath, sequence, and collision-free motion planning that is orders of magnitude faste

desk verdict Solid, well-engineered pipeline with real physical prints, but the 'strictly collision-free' claim is about an idealized model of deposition, not the actual extruded bead. read the letter →

arxiv 2509.05345 v1 pith:FGFILTXN submitted 2025-09-02 cs.RO cs.CG

Jiasheng Qu , Zhuo Huang , Dezhao Guo , Hailin Sun , Aoran Lyu , Chengkai Dai , Yeung Yam , Guoxin Fang This is my paper

classification cs.ROcs.CG

keywords implicitneuralfieldsmulti-axis3Dprintingcollision-freemotionplanningtoolpathgenerationsigneddistancefieldtime-varyingSDFquaternionoptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the bottleneck in multi-axis 3D printing is representational: explicit meshes, voxels, and tetrahedra make toolpath generation resolution-dependent and global collision checking prohibitively slow as the printed object grows. The authors propose to encode the entire fabrication process as implicit neural fields—the model geometry as a signed distance field, the print direction as a guidance field, the infill as separate fields, the printing order as a sequence field, and the robot motion as a quaternion field—so that every stage is continuous, differentiable, and GPU-parallel. From these fields, shell and infill toolpaths are extracted by intersecting iso-surfaces, and the evolving printed object is modeled as a time-varying SDF that supports exact-sign collision evaluation during motion optimization. The paper reports that this removes resolution dependence, gives waypoint errors as low as 0.019 mm, reduces computation time by up to two orders of magnitude relative to mesh-based slicing, and produces physical prints with fewer defects. If correct, the significance is that the hardest scaling problem in multi-axis printing—planning millions of collision-free waypoints—becomes a field optimization that does not degrade with model size.

What carries the argument

The carrying object is the time-varying signed distance field φ_T(p, x), built by interpolating the static model SDF φ and the printing sequence field T. It converts the collision-checking problem—normally a per-timestep geometric intersection test against a growing mesh—into a pointwise sign and distance evaluation that is differentiable with respect to the motion quaternion field q. Supporting it are the guidance field g, whose gradient defines deposition direction; infill fields ψ; and a quaternion field q for extruder orientation. Waypoints are produced by projecting grid candidates onto the intersection manifold g = u, ψ = v, and the Reeb-graph-derived partition ℓ keeps the sequence con

What would settle it

Print a curved thin-wall or overhang model, pause mid-print at several waypoints, and 3D-scan the partial object; compare the scanned geometry against the sublevel set predicted by the sequence field and against the time-varying SDF's collision contours. If any scanned cross-section shows deposited material outside the predicted sublevel set (i.e., bead overfill) that intersects the extruder assembly in a configuration the planner declared collision-free, the central guarantee is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a single family of implicit neural fields can carry all the information needed to print an object on a multi-axis robot: the SDF encodes the target geometry; a learned guidance field encodes a smooth, singularity-controlled print direction that is support-free on the surface and manufacturable inside; infill fields encode controllable, support-free internal structure; a sequence field orders the deposition; and a quaternion field orients the extruder. The key construction is the time-varying SDF: rather than training a network that takes time as input, the paper interpolates between the final-model SDF and the sequence field, so the set of already-printed points is

Load-bearing premise

The printed-object geometry at each moment is assumed to be exactly the set of points whose sequence value is already reached and whose signed distance is non-positive, ignoring how wide the extruded bead is and the path the nozzle sweeps.

Editorial extensions

If this is right

Toolpath generation for shell and infill becomes a single field-optimization problem, so models of arbitrary input format (point clouds, B-rep, meshes) share the same pipeline.
Global collision avoidance becomes a differentiable loss, so printing sequence and robot motion can be co-optimized and iteratively refined until the collision rate reaches zero for all waypoints.
Waypoint generation scales to millions of points in seconds on a GPU, because discretization is done by parallel projection rather than sequential tracing along meshes.
Physical prints show support-free fabrication with measured distance error reduced by about 28–29% and visual normal error by 39–78% compared with a mesh-based slicer.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The time-varying SDF construction is a general trick for any process with a monotonically growing workpiece—subtractive manufacturing access analysis, welding, or robotic assembly—where the 'already processed' set is defined by a scalar sequence field.
The fidelity of the collision-free guarantee depends on how well the sublevel-set model matches the real bead cross-section; a plausible failure mode is that thin-wall or high-curvature regions where the bead overfills the modeled sublevel set will show collisions the planner did not see. A test would be scanning a partially printed model mid-process.
Because the distance approximation in Eq. 27 is a local linearization, the gradient of the collision loss may become inaccurate for points far from the current print surface; the method's practical robustness could be probed by stress-testing on models whose waypoint density is low relative to the robot's size.
The claimed 'first systematic pipeline' status is a scope claim about combining shell and solid printing with global collision avoidance; the more consequential legacy may be showing that a single neural field can replace the discrete pipeline wholesale, which can be tested on other process planning problems.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a solid, well-engineered paper that deserves a real referee and likely publication after revision, but the headline 'strictly collision-free' is stronger than what is actually shown.

What is new: the full implicit pipeline—SIREN SDF, guidance field, infill fields, sequence field, quaternion field—and the training-free time-varying SDF built by field interpolation. That removes explicit volumetric meshes from the loop, which is a genuine step beyond Neural-Slicer and S3-slicer. The waypoint generation via Marching-Cubes initialization plus field projection is a clean, parallelizable solution, and the reported waypoint errors are much lower than the mesh-based baselines. Physical fabrication of five models, including scans against S3-slicer, is real evidence. The ablations on two-stage guidance training and on separating LPD/self-rotation smoothness are honest and useful.

Soft spots, in order of size. First, the collision-free guarantee is for an idealized model of deposition, not the physical bead. Algorithm 1 treats the printed region as the sublevel set {phi<=0 and T<=T(p)}. That ignores extrusion width and toolpath sweep; material can be deposited across g iso-surfaces into regions with T>T(p), and line 7 skips those points. So the collision checker can call a waypoint safe while the physical nozzle is already inside deposited material. The paper's own caveat in 6.6.2 that only the last two joints plus extruder are modeled, as point clouds, compounds the issue. Second, the collision rate used as a success metric is computed by the same time-varying SDF that the optimization uses, so a zero collision rate is partly self-consistent. Independent support comes from the physical prints, but those validate selected trajectories, not every waypoint against measured deposition. Third, the convergence guarantee for iterative partition refinement is asserted, not proven: 'worst case T=g' does not by itself imply collision-free. Fourth, the open-source link lacks a commit hash and reproduction instructions.

None of this sinks the paper. The practical claims are demonstrated on the tested models, and the physical prints succeeded. But the authors should reword the guarantees, separate the success metric from the predicate being optimized, and either include a bead-aware check or state the limitation explicitly. I would send it to peer review and push for major/minor revision along those lines. Worth discussing in reading group and citing for the pipeline itself.

Referee Report

4 major / 5 minor

Summary. INF-3DP presents a fully implicit-neural-field pipeline for multi-axis 3D printing. A SIREN-based SDF represents the input model; a guidance field is trained in two stages over surface and interior to enforce support-free, smooth, singularity-aware toolpaths; infill fields produce regular or variable-density patterns; the printing sequence is encoded by a Reeb-graph-partitioned field T = g + ℓ; and a quaternion field q encodes tool orientation. The dynamically growing object is represented by a time-varying SDF φ_T built by field interpolation from φ and T, enabling a differentiable collision loss for motion optimization. Waypoints are generated by parallelized field projection onto implicitly defined iso-curves. Validation includes ablations, comparisons to S3-slicer and Neural-Slicer, scalability statistics, and physical prints of five models on a UR5e robot.

Significance. If the stated guarantees hold, this is a substantial advance: it is, to my knowledge, the first multi-axis 3D-printing pipeline that uses INFs for all stages from geometry through toolpath generation to motion planning, and the reported waypoint-accuracy improvements and order-of-magnitude speedups are credible and practically useful. The paper ships an open-source implementation, detailed per-stage timings, ablations, and physical fabrication evidence, all of which are strengths. The central limitation is that the collision-free guarantee is model-relative: the evolving object is an idealized sublevel-set truncation of the final SDF, the robot is reduced to the last two joints plus extruder, and the collision metric is evaluated with the same time-varying SDF used for optimization. The headline 'strict/global collision-free' claims therefore need careful scoping.

major comments (4)

[Sec. 5.1, Algorithm 1 (lines 6–8), Eq. (27)] The evolving printed object is modeled as {x | φ(x)≤0 and T(x)≤T(p)}. In a filament process, the deposited bead has finite width and is swept along the toolpath, so material can be present at points with T(x)>T(p) and even φ(x)>0. Algorithm 1 line 7 skips exactly those points as collision-free. Thus the 'exact collision checking' statement after Eq. (27) is exact only for the ideal deposition model, not for the physical object. Since 'strictly collision-free' is a headline contribution (Sec. 1.1, Sec. 6.1.3), please either incorporate a deposition-aware occupancy model or explicitly restrict the claim and provide an independent check of the collision predicate against measured or simulated swept material.
[Eqs. (23), (27); Sec. 5.3] The collision loss is L_coll = Σ −φ_T, with φ_T built using the first-order approximation d ≈ (g_k − g(x))/||∇g(x)||. The paper argues this approximation does not affect collision checking because signs come from φ and T, but the magnitude of φ_T and the updates in Algorithm 1 depend on the approximation. Moreover, the 'exact collision ratio' used for early stopping is evaluated with this same time-varying SDF. A zero collision rate is therefore partly a self-consistency check. I recommend an external validation: for a representative set of waypoints, reconstruct the swept filament volume from the planned path and check it against a discretized model of the end-effector, then report the resulting collision rate.
[Sec. 4.4 and Sec. 6.6.2] The statement that iterative partition refinement is guaranteed to converge—'in the worst case, the process converges to T(x)=g(x)'—is asserted without proof and assumes that a collision-free, support-free, smooth motion exists in the restricted task space. The acknowledged simplification in Sec. 6.6.2 (only last two joints plus extruder, M_e discretized as point clouds) means that collisions of other robot links and between sample points are not checked. As written, the Table 1 row 'Global Collision-Free: Yes' is stronger than what is established. Please restate the guarantee as model-based and task-space-specific, and either provide a formal convergence argument or present the iteration as an empirically convergent heuristic.
[Sec. 5.1, Algorithm 1 line 12] The text defines g_k = max{g(x) | ℓ(x)=ℓ_k ∧ T(x)≤T(p)}, but the pseudocode line 12 computes g_k = max{g(x) | ℓ(x)=ℓ_k} without the T≤T(p) restriction. For ℓ_k = ℓ(p), this includes points that have not yet been printed at waypoint p, so the top surface Stop(ℓ_k) can lie above the current print front and distort the distance approximation in Eq. (27). Please align the pseudocode with the text and fix the undefined `T(g)` in the same paragraph.

minor comments (5)

[Eq. (2)] The notation d_p(x) is not defined; presumably it is the unit local printing direction. Please define it explicitly. Also, the printed constraint is written with an absolute value but the subsequent support-free losses use a sigmoid relaxation; make the relationship between C_SF and L_SF precise.
[Fig. 9 and Sec. 6.3] The text refers to 'right of Fig. 9(e)', but the figure caption lists panels (a)–(d) and no (e). Please correct the cross-reference or add the missing panel.
[Algorithm 1, line 13] Typo: 'Approximiate' should be 'Approximate'. In Sec. 5.1, 'closet top surface' should be 'closest top surface'. These should be fixed before publication.
[Sec. 6.5] Physical printing parameters are incomplete: layer height, feed rate, extrusion temperature, and robot speed are not reported, which makes the physical validation difficult to reproduce. At least the key parameters should be added.
[Conclusion] The phrase 'collision-free, high surface finishing, and collision-free fabrication' repeats 'collision-free'; the first instance should likely be 'support-free' or the sentence should be rephrased.

Circularity Check

1 steps flagged · score 3.0 of 10

One mild self-consistency: the reported zero collision rate is the training objective's stopping condition; physical fabrication and external baselines provide independent support.

self definitional [Sec. 3.1.3 (Eq. 6), Sec. 4.4 (Eqs. 22-23), Sec. 5.3, Sec. 6.1.3]
"the collision loss Lcoll is computed as the sum of the distances from given points to the surface of the partially printed object at each sequence step T(p). Specifically, it is evaluated as −φT(p, x), as detailed in Sec. 5.1. // Early stopping is triggered only when the collision ratio converges to zero, ensuring a restricted global collision-free motion sequence is found."

The motion quaternion field q is optimized by minimizing Lcoll = Σ_{p,x} −φT(p,x) over points where φT<0, while the reported success metric is Ccoll = ∫ H(−φT) ≡ 0 over all waypoints. Thus the 'collision-free' result is exactly the optimizer's stopping condition evaluated on the same time-varying SDF that defines the collision predicate; it holds by construction for the model, not as an independent check. Physical fabrication gives external evidence, but it does not measure the collision predicate at every waypoint against actual deposited material, so the model-level guarantee remains self-consistent rather than independently confirmed.

full rationale

The central derivation chain is otherwise self-contained. No load-bearing self-citation or imported uniqueness theorem appears; the Reeb-graph construction cites external work (Zhong et al. 2023a) and the TV-SDF interpolation cites Marschner et al. 2023. The absolute waypoint-accuracy number is partly the residual of the projection step in Eq. 30, but the reported advantage over mesh-based baselines (S3-slicer, Neural-Slicer) is an external comparison, and the physical fabrication of five models independently confirms executability. The paper also scopes its collision model to the last two joints plus extruder (Sec. 6.6.2), which is a limitation rather than a circular step. Overall, one metric/objective coincidence warrants a score of 3, not higher, because the main claims retain independent empirical content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a set of modeling choices: the truncated-SDF representation of the growing object, the approximated distance in Eq. 27, the simplified robot model, and an unproven convergence statement for the iterative partition refinement. These are not new physical entities but computational assumptions. The many loss weights and thresholds are tuned by hand; they influence the results but are not presented as predictions.

free parameters (6)

Loss weighting coefficients (omega_dist, omega_norm, omega_NM, omega_eikonal, omega_lap, omega_S_smooth, omega_S_SF, ome = not reported in text
Relative weights of fabrication objectives in Eqs. 8, 9, 13, 17, and 22 are chosen by hand; the paper does not give values or a sensitivity study. The balance between support-free, smoothness, and collision objectives determines the quality of the final plans.
Support-free angle alpha = 40 degrees for PLA
Material-dependent threshold used in Eqs. 2, 11, 20, and 24. It defines what counts as support-free and is taken as an input from prior literature for PLA.
Sigmoid sharpness k_SF = 30
Relaxation parameter for support-free losses in Eqs. 11, 20, and 24, set by hand.
Singularity detection threshold epsilon = 1e-2
Norm of the tangent vector field below which a singularity is declared (Sec. 3.1.2). It influences where singularities are allowed and affects surface quality optimization.
Reeb-graph neighbor threshold = 2 times average layer height
Threshold for connecting iso-contour region centers into the Reeb graph (Sec. 4.4), directly controls partitioning and printing continuity.
Non-manifold decay gamma = much greater than 1
Used in the SDF non-manifold loss (Eq. 8) to regularize regions far from the surface; the specific value is not given.

assumptions (6)

domain assumption The printed object at waypoint p is the final model SDF truncated by the printing sequence field: {x | phi(x) <= 0 and T(x) <= T(p)}
Invoked in Sec. 5.1 and Algorithm 1 to construct the time-varying SDF. It ignores extrusion width, actual deposition along toolpaths, and the fact that not every point with a smaller sequence value is filled at the moment waypoint p is printed.
domain assumption Point-to-top-surface distance can be approximated by (g_k - g(x)) / ||grad g(x)|| (Eq. 27)
Used to update the time-varying SDF and compute collision loss. The approximation is accurate near the current surface but degrades far away; the paper argues this does not affect collision identification because sign checks use phi(x) and T(x).
ad hoc to paper The printing setup can be represented by the point cloud of the last two joints plus the extruder, and the robot base never collides
Sec. 6.6.2 explicitly limits the setup to the final two joints and extruder, ignoring full inverse kinematics and the base. The authors state no base collisions occurred in their tests, but this is not guaranteed in general.
ad hoc to paper The iterative partition refinement converges to a collision-free solution
Sec. 4.4 states convergence is guaranteed, with worst-case T(x)=g(x), but no proof is provided. The claim that the process always finds a feasible partition is asserted without derivation.
domain assumption SIREN networks represent the fields with sufficient accuracy for all tested models
The pipeline depends on the trained INFs (SDF, guidance, infill, sequence, quaternion) being accurate enough for waypoint projection and collision checking. The paper reports average SDF error below 2.5e-4 but no per-model error bounds for the other fields.
domain assumption Support-free condition with alpha=40 degrees is necessary and sufficient for printability
Uses the cosine of the angle between the local printing direction and the surface normal; physically, material can be deposited without support if this angle is below a threshold. Standard in the literature; taken as given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of INF-3DP: Implicit Neural Fields for Collision-Free Multi-Axis 3D Printing." pith.science (2026). https://pith.science/paper/FGFILTXN

@misc{pith2026250905345,
  author       = {Pith},
  title        = {Pith review of: INF-3DP: Implicit Neural Fields for Collision-Free Multi-Axis 3D Printing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FGFILTXN}},
  note         = {Machine review of arXiv:2509.05345}
}

read the original abstract

We introduce a general, scalable computational framework for multi-axis 3D printing based on implicit neural fields (INFs) that unifies all stages of toolpath generation and global collision-free motion planning. In our pipeline, input models are represented as signed distance fields, with fabrication objectives such as support-free printing, surface finish quality, and extrusion control being directly encoded in the optimization of an implicit guidance field. This unified approach enables toolpath optimization across both surface and interior domains, allowing shell and infill paths to be generated via implicit field interpolation. The printing sequence and multi-axis motion are then jointly optimized over a continuous quaternion field. Our continuous formulation constructs the evolving printing object as a time-varying SDF, supporting differentiable global collision handling throughout INF-based motion planning. Compared to explicit-representation-based methods, INF-3DP achieves up to two orders of magnitude speedup and significantly reduces waypoint-to-surface error. We validate our framework on diverse, complex models and demonstrate its efficiency with physical fabrication experiments using a robot-assisted multi-axis system.

Figures

Figures reproduced from arXiv: 2509.05345 by the authors.

**Figure 1.** We present INF-3DP, a computational framework leveraging Implicit Neural Fields (INFs) to optimize toolpath generation and plan global collision-free [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Fabrication objectives for multi-axis 3DP are closely related to field [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The computational pipeline of INF-3DP. (a) Begins with representing the printing model [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: (a) For the Satellite Bracket model with complex topology, initial partitioning with a Reeb graph maximizes print continuity, but directly using [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Illustration of the time-varying SDF during material deposition: (a) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Different types of infill patterns can be achieved by efficiently train [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: Computation results and fabrication process for the statue-goddess [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Computational results for toolpath generation and motion planning on the Light Bulb model. (a) SDF trained using vertices from a dense mesh (average [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Training process and ablation study for the guidance field on the Fertility model. (a) Visualization of the vector field smoothness [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 11.** Figure 11: Result of physical fabrication. From left to right: Statue-Goddess, [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: Comparison of fabrication errors with toolpaths generated by the baseline method [Zhang et al [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

**Figure 13.** Figure 13: Comparison of performance with Neural-Slicer [Liu et al. 2024] on the spiral fish model. (a) Both methods generate support-free toolpaths, but our method (b) achieves superior surface finishing by optimizing singularity placements. (c) With an input model represented …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

[2015]

DeepMill: Neural Accessibility Learning for Subtractive Manufacturing

Perceptual models of preference in 3D printing direction. ACM transactions on graphics (TOG) 34, 6 (2015), 1–12. Haisen Zhao, Hao Zhang, Shiqing Xin, Yuanmin Deng, Changhe Tu, Wenping Wang, Daniel Cohen-Or, and Baoquan Chen. 2018. DSCarver: decompose-and-spiral-carve for subtractive manufacturing. ACM Transactions on Graphics (TOG) 37, 4 (2018), 1–14. Fan...

work page Pith review arXiv 2015
[2019]

ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–11

CurviSlicer: Slightly curved slicing for 3-axis printers. ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–11. Guoxin Fang, Tianyu Zhang, Yuming Huang, Zhizhou Zhang, Kunal Masania, and Charlie CL Wang. 2024. Exceptional mechanical performance by spatial printing with continuous fiber: Curved slicing, toolpath generation and physical verification. Addit...

arXiv 2019
[2020]

NeurCross: A Neural Approach to Computing Cross Fields for Quad Mesh Generation

Planning jerk-optimized trajectory with discrete time constraints for redun- dant robots. IEEE Transactions on Automation Science and Engineering 17, 4 (2020), 1711–1724. Chengkai Dai, Charlie CL Wang, Chenming Wu, Sylvain Lefebvre, Guoxin Fang, and Yong-Jin Liu. 2018. Support-free volume printing by multi-axis motion. ACM Transactions on Graphics (TOG) 3...

work page Pith review arXiv 2020
[2024]

Journal of Computing and Information Science in Engineering 24, 5 (2024)

Vector Field-Based Volume Peeling for Multi-Axis Machining. Journal of Computing and Information Science in Engineering 24, 5 (2024). Jimmy Etienne, Nicolas Ray, Daniele Panozzo, Samuel Hornus, Charlie CL Wang, Jonàs Martínez, Sara McMains, Marc Alexa, Brian Wyvill, and Sylvain Lefebvre

work page 2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [2015]

DeepMill: Neural Accessibility Learning for Subtractive Manufacturing

Perceptual models of preference in 3D printing direction. ACM transactions on graphics (TOG) 34, 6 (2015), 1–12. Haisen Zhao, Hao Zhang, Shiqing Xin, Yuanmin Deng, Changhe Tu, Wenping Wang, Daniel Cohen-Or, and Baoquan Chen. 2018. DSCarver: decompose-and-spiral-carve for subtractive manufacturing. ACM Transactions on Graphics (TOG) 37, 4 (2018), 1–14. Fan...

work page Pith review arXiv 2015

[2] [2019]

ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–11

CurviSlicer: Slightly curved slicing for 3-axis printers. ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–11. Guoxin Fang, Tianyu Zhang, Yuming Huang, Zhizhou Zhang, Kunal Masania, and Charlie CL Wang. 2024. Exceptional mechanical performance by spatial printing with continuous fiber: Curved slicing, toolpath generation and physical verification. Addit...

arXiv 2019

[3] [2020]

NeurCross: A Neural Approach to Computing Cross Fields for Quad Mesh Generation

Planning jerk-optimized trajectory with discrete time constraints for redun- dant robots. IEEE Transactions on Automation Science and Engineering 17, 4 (2020), 1711–1724. Chengkai Dai, Charlie CL Wang, Chenming Wu, Sylvain Lefebvre, Guoxin Fang, and Yong-Jin Liu. 2018. Support-free volume printing by multi-axis motion. ACM Transactions on Graphics (TOG) 3...

work page Pith review arXiv 2020

[4] [2024]

Journal of Computing and Information Science in Engineering 24, 5 (2024)

Vector Field-Based Volume Peeling for Multi-Axis Machining. Journal of Computing and Information Science in Engineering 24, 5 (2024). Jimmy Etienne, Nicolas Ray, Daniele Panozzo, Samuel Hornus, Charlie CL Wang, Jonàs Martínez, Sara McMains, Marc Alexa, Brian Wyvill, and Sylvain Lefebvre

work page 2024