WorkBenchMark: Planning LEGO Duplo Assembly by Disassembly
WorkBenchMark tests whether a robot can assemble LEGO Duplo products from loose parts. Its baseline combines observed geometry, an assembly order inferred by taking the target apart, and precise motion. This robotics benchmark is separate from BrickGPT.online.
A benchmark that includes physical work
The June 2026 arXiv v2 paper defines 400 simulation tasks, 100 in each of four tiers: basic vertical stacking, multi-brick sequences, spatial layouts and complex interlocking structures. A task supplies a target product and scattered parts. The robot must identify parts, order placements, grasp and insert them without collisions or collapse.
A correct abstract order can still fail because the gripper cannot reach a piece or a later press destabilizes the assembly. The paper also evaluates a smaller real-robot subset.
Infer assembly order by taking the target apart
Assembly-by-Disassembly represents the target as an occupied voxel state and recursively removes top-down candidates. Each candidate needs a collision-free grasp pose, and the remaining assembly must support future pressing. The latter is approximated by checking a projected press point against a support polygon formed by ground-contacting bricks.
The search backtracks if a branch fails and can handle connected subassemblies together. Reversing a successful disassembly sequence produces the assembly order. The support-polygon check is an approximation, not a full contact-physics proof.
Ground parts visually, then execute collision-aware motions
GroundingDINO locates parts from text descriptions, SAM refines instance masks, and FoundationPose estimates six-degree-of-freedom poses using known CAD meshes. The controller follows approach, open, descend, grasp, lift, move, descend and release.
MoveIt 2 and OMPL plan collision-free transit motions; constrained Cartesian descents align grasp and insertion. If execution fails, the system tries another grasp and may replan. Known meshes, camera calibration and reliable pose estimates remain important dependencies.
What the reported results show
Across 400 simulation tasks, the structured pipeline reports 77.5% overall success, compared with 42.5% for a fine-tuned VLA baseline and 38.25% for a zero-shot VLM/VLA baseline. Its success falls from 94% in Tier 1 to 62% in Tier 4. On a UR5, the paper reports 90%, 90% and 70% success on 10 real-robot tasks per tier for Tiers 1-3; Tier 4 was not included in that physical test.
These measurements belong to the defined parts, simulator and robot. They do not establish performance on arbitrary sets or imply that BrickGPT.online can drive a robot.
Read the original research
The study separates three failures a visual assembly story can conceal: incorrect part pose, impossible placement order and failed physical insertion. Consult the paper for its exact task constraints and protocol.
BrickGPT Lab