The open-source 14-DoF DIYRobot platform

Kinematics-supervised expert routing

KinRT

Route by Kinematics, Act by Observation

Explicit kinematics-guided MoE routing distills action priors into observation space.

Tianhang Yang*, Yanze Zheng*, Junjie Wang, Wei-Bin Kou, Ruotong Li, Yujiu Yang

Tsinghua University · Pengcheng Lab · *Equal contribution · Corresponding authors

The routing problem

Visual similarity is not motion similarity.

Implicit MoE routers learn expert assignment from appearance and language alone. Robot manipulation breaks that assumption: similar scenes can demand different trajectories, while different objects can share the same kinematic structure.

KinRT uses privileged action information only during training. It clusters action and velocity into kinematic archetypes, supervises the router with those labels, and transfers that structure into observation-only routing at inference.

4kinematic archetypes
Top-1global routing
50steps per action chunk
0.05router loss coefficient

Method

Supervise with actions. Route with observations.

KinRT discovers motion archetypes offline, uses their IDs to train a lightweight router, and activates one routed expert alongside a shared expert for every action chunk.

KinRT clustering, global router, action-expert pipeline, and DIYRobot platform
Action and velocity form the training-only supervisory signal. The deployed router receives observation embeddings and never requires future actions.
01

Discover archetypes

Encode 50-step trajectories, apply PCA, and cluster the action space into coherent motion regimes.

02

Teach the router

Use cluster IDs as explicit labels for a global router trained from pooled observation embeddings.

03

Deploy without actions

Select one specialized expert using the current images, instruction, and proprioceptive state.

Key results

Consistent gains across simulation and hardware.

Reported values are average successful trials across five DIYRobot tasks and eight RoboTwin tasks.

+20.27%DIYRobot
+23.26%RoboTwin Easy
+13.78%RoboTwin Hard
BenchmarkDense baselineKinRTAbsolute gainRelative gain
DIYRobot29.635.6+6.0+20.27%
RoboTwin 2.0 Easy / Hard33.1 / 34.140.8 / 38.8+7.7 / +4.7+23.26% / +13.78%

KinRT also surpasses the strongest implicit-routing MoE baseline, AdaMoE, by +14.2 successes on DIYRobot and by +8.7/+9.4 successes on RoboTwin Easy/Hard.

Real-robot demonstrations

DIYRobot in action.

Representative physical rollouts show KinRT controlling coordinated, contact-rich tasks from three complementary viewpoints. Residual action jitter can be mitigated through output smoothing, temporal ensembling, and controller-level tuning.

Dual-arm handoverCoordinated transfer between end effectors
Press buttonTargeted contact with visual feedback
Rotate screwdriverGrasp stabilization and wrist rotation

Evaluation

Two platforms, one controlled comparison.

DIYRobot keeps a 50-trial original-environment test protocol, with altered-lighting demonstrations available only as optional training data. RoboTwin uses Easy and Hard conditions.

Open-source reproduction

From source audit to robot deployment.

The release keeps KinRT FULL and LoRA in one source tree and separates simulator, dataset, checkpoint, and hardware prerequisites explicitly. Start with the software-only validation gate before training or connecting a robot.

Citation

Cite KinRT

Use the arXiv record while the archival publication metadata is pending.

BibTeX
@article{yang2026route,
  title={Route by Kinematics, Act by Observation:
    Kinematics-Supervised Expert Routing in MoE-Augmented VLA},
  author={Yang, Tianhang and Zheng, Yanze and Wang, Junjie and
    Kou, Wei-Bin and Li, Ruotong and Yang, Yujiu},
  journal={arXiv preprint arXiv:2607.26807},
  year={2026}
}