MARC: Mechanism-AwareReflective Correctionfor Articulated Object Code Generation

Xiong Yang1,2,*Jinkun Hao1,*Zijian Xiao1,2Zipeng Ye2,†Yuchen Xie2Ran Yi1,†,‡
1 Shanghai Jiao Tong University2 Meituan

* Equal contribution   † Project leader   ‡ Corresponding author

Paper Video Visual supplement
A MARC-generated scissor lift extending and contractingScissor lift
A MARC-generated piston engine with coordinated crankshaft and piston motionPiston engine
A MARC-generated planetary gearset in motionPlanetary gearset
From images and text to articulated programs with coordinated motion.

One kinematic graph connects how a mechanism is built, how it moves,
and how the agent repairs it.

Abstract

Generating an articulated object from images and text requires more than plausible parts and movable joints: the assembled mechanism must execute the intended motion. We introduce MARC, Mechanism-Aware Reflective Correction, an agent framework that shares an executable kinematic graph across program construction, motion solving, and correction. Physical connections are bound to local attachment frames, while loop closures and joint-coordinate couplings describe how the parts move together. A numerical solver computes follower motion and associates anomalies with specific connections and configurations. The agent uses this information to probe possible causes, revise connections or geometry, and test the mechanism again. On a 20-object test set, MARC with Kimi-K3 achieves 67.8% valid motion coverage and 60.0% overall success. Starting from the same initial candidates, reflective correction improves these scores by 13.7 and 20.0 percentage points, respectively.

Video overview

Construction, motion solving, and correction in 90 seconds.

MARC · Project overviewDownload MP4 ↓

A shared graph for generation and repair

Motion errors point back to the connections and geometry the agent can edit.

MARC represents rigid bodies, physical connections, loop closures, and coordinate couplings in one executable kinematic graph. Each connection is bound to local attachment frames, or sockets, so its geometric location stays explicit throughout construction, numerical execution, and correction.

MARC workflow: Topology builds the mechanism, Solver computes motion, Evaluate checks the result, and Judge, Executor, and Reviewer guide correction
MARC's generation and correction workflow. The socket edit in the diagram is schematic. View full size ↗
1

Construct the mechanism

The agent writes a program defining geometry, socket-bound connections, and motion relationships through a mechanism-authoring SDK.

2

Solve and inspect motion

A numerical solver computes follower motion from prescribed driver inputs, recording anomalies at the bodies, joints, and sockets involved.

3

Locate, revise, and retest

The agent probes possible causes, coordinates edits to connections and geometry, and reruns the complete mechanism along the original motion paths.

Look at how the parts move together

Connection and coupling behavior in a generated dual-head lamp.

Keeping the linkage connected

As the lamp extends, its crossing links and pivots must stay connected. Compare how the generated assemblies respond to their control inputs.

Input reference for the dual-head lamp
Input reference
ArticraftKimi-K3
Articraft dual-head lamp at a recorded motion state
Native control
MARC OursKimi-K3
MARC dual-head lamp at a recorded motion state
Native control
0%

Separate outputs with the same backbone. Recorded native-control states are aligned by sequence progress.

Compare all seven methods →

Reflective correction improves valid motion

From the same initial candidates, more mechanisms complete their prescribed motion.

20 objects · Kimi-K3 · Same initial candidates · Paper, Table 4

Valid motion coverage

54.1%→67.8%
Without reflection
Full MARC

+13.7 percentage points

Overall success rate

40.0%→60.0%
Without reflection
Full MARC

+20.0 percentage points

Valid motion coverage measures the proportion of prescribed samples reached while satisfying connection, joint-limit, and collision criteria. Overall success additionally requires the critical structural and motion requirements and complete valid sampled coverage.

Full comparisons and ablations in the paper →

Citation

Preprint citation. The arXiv identifier and link will be added when available.

@misc{yang2026marc,
  title  = {{MARC}: Mechanism-Aware Reflective Correction
            for Articulated Object Code Generation},
  author = {Xiong Yang and
            Jinkun Hao and
            Zijian Xiao and
            Zipeng Ye and
            Yuchen Xie and
            Ran Yi},
  year   = {2026},
  note   = {Preprint}
}