A Modular Path Planner for Robotic Manufacturing That Actually Ran on Real Hardware

Offline optimization planner for robot arms in additive manufacturing, validated on a UR10e with carbon-fiber winding and 3D printing. The modularity story is real; the validation is thin.

Imagine you're planning a road trip with multiple constraints: you need to hit certain landmarks in order, stay under a fuel budget, avoid toll roads, and keep the passengers comfortable. You could plan the whole route at once — but if it's a cross-country trip, no GPS can handle the full problem without choking. So you split it into day-legs, stitch the legs together at overnight stops, and iteratively adjust the budget tradeoffs until every day meets its targets. That's essentially what this paper does for robot arm paths in additive manufacturing. The committed claim: a single offline optimization framework that can absorb different manufacturing processes — carbon-fiber winding, robotic 3D printing — by swapping in process-specific cost terms and constraints, without redesigning the planner. The paper's structural contribution is that modularity layer, plus an Iterative Weight Tuning (IWT) scheme that auto-tunes cost weights until Key Performance Indicators hit their targets. It's an engineering integration result, not a fundamental algorithmic breakthrough. The architecture sits squarely in nonlinear programming territory — joint-space trajectory optimization with quadratic regularization, self-collision penalties, and process-specific residuals. The chunked optimization approach (splitting long paths into boundary-coupled segments) is a well-known decomposition strategy, but the specific coupling of chunks with shared boundary conditions and the IWT loop on top is a clean engineering contribution. The solver runs offline, meaning this is not competing with real-time MPC approaches — it's a pre-planning tool. On the ladder, this is harder to assess than it should be. The abstract names no competing method, no baseline planner, and no quantitative comparison against existing path planners like CHOMP, TrajOpt, STOMP, or any MoveIt-based pipeline. The validation is: planned paths met geometric tolerances and KPI targets, and the paths executed on a UR10e with joint-level tracking recorded. That's a necessary demonstration but not a comparative one. We know it worked; we don't know if it worked better than alternatives. Integrity is the weakest dimension. The validation is same-team, same-hardware, no independent replication, no community benchmark, and the printing path was executed without extrusion — meaning the actual material deposition quality was not tested. Joint-level tracking during execution is recorded but not compared against, say, a standard trajectory planner's tracking performance on the same geometry. The carbon-fiber winding did include extrusion, which is stronger, but quantitative results are not in the abstract. The milestone question for this subfield is convergence toward closed-loop, online-adaptive planners that handle process disturbances in real time. This paper is firmly in the offline camp. The next concrete step would be demonstrating the framework on a longer, more complex geometry where the chunked optimization's scalability advantage becomes measurable — say, a full aerospace composite layup with 500+ waypoints — and comparing wall-clock planning time and path quality against TrajOpt or CHOMP on the same problem. The obvious experiment not run: full 3D printing with actual extrusion and material quality assessment. The abstract states printing was executed "without extrusion." The honest read is (a) — the experimental setup for in-process extrusion quality monitoring was likely not available or not within scope for a conference submission targeting ECC 2027. This is a controls conference paper, not a manufacturing science paper, so the omission is understandable but limits the practical claim significantly.