MEKAWIRE
ExpertarXiv

PolyMorphWBC: Whole-Body Control from End-Effector Trajectories for Quadrupedal Mobile Manipulation across Arm Morphologies with Extended Reach

Heng Li, Yang Li, Yifan Duan, Ren Fang, Matao Duan, Wenhao Yu, Lu Zhang, Jianmin Ji, Yanyong Zhang
In 30 seconds

This preprint introduces PolyMorphWBC, a controller that coordinates a quadruped and its mounted arm using only end-effector trajectories as task-level motion commands. Across held-out arm morphologies and trajectories in IsaacLab, it achieved 93.95% trajectory completion with 2.30 cm position error and 4.81° orientation error. The approach could simplify connecting manipulation policies to mobile robots while reusing a controller across different arms.

Research question

Can one control hierarchy turn end-effector trajectories into coordinated arm and base motion across different mounted-arm morphologies, including targets beyond initial fixed-base reach?

Problem

An end-effector trajectory leaves the controller to decide when and how the base should move. Waiting too long makes upcoming targets unreachable, while unnecessary movement can worsen tracking. Different arm geometries, mass distributions and motion-induced disturbances change the required response.

Previous approach

Optimization-based systems such as ALMA and unified MPC coordinate locomotion and manipulation through models. Learned approaches include DeepWBC, VBC and RoboDuet. Some interfaces require base commands, while UMI-on-Legs and MLM accept end-effector trajectories but were developed around fixed mounted-arm morphologies.

New approach

A shared learned coordinator previews 7 end-effector poses spanning 1 s and selects base velocity, yaw rate, pitch and height commands. Arm-specific inverse kinematics tracks the current target, and a frozen locomotion policy controls the legs. A URDF-derived morphology encoder, pretrained on kinematics and residual-wrench prediction, supplies static arm information; estimated residual wrenches and recent response history supply online feedback. Staged PPO training uses 1024 training arms, 128 validation arms and 100k training trajectories across four 6-DoF arm families. For hardware pick-and-place, π₀ was fine-tuned with 100 demonstrations and supplied end-effector trajectories and gripper commands.

Results

IsaacLab held-out evaluation achieved 93.95% completion, 2.30 cm mean position error and 4.81° mean orientation error. For trajectories containing x>1 m targets in the initial base frame, results were 94.83%, 1.94 cm and 3.86°; for x<0 m, 91.56%, 2.41 cm and 4.68°. Errors include failed episodes. Held-out banks contain 128 arms and 10k trajectories; the total rollout count is not reported. On four Go2-ARX5 trajectories in MuJoCo without retraining, position errors were 1.38 to 2.04 cm and orientation errors 2.01° to 2.38°, versus 4.77 to 7.31 cm and 22.72° to 76.57° for the official UMI-on-Legs random-trajectory checkpoint without task fine-tuning. On horizontal base orbit specifically: 2.04 cm and 2.38° versus 7.31 cm and 76.57°. Full-model completion minus completion without morphology conditioning was +9.90 percentage points for CW, +2.59 for OW, +0.55 for SW and -0.48 for PA. Physical Go2-ARX5 demonstrations showed teleoperation and language-conditioned extended-reach pick-and-place without task-level base commands.

Limitations

The method assumes known arm models, smooth low-speed trajectories and a common Go2 base. Cross-arm evidence is simulated; hardware demonstrations use one arm configuration. Deployment envelopes are empirical operating limits, not feasibility guarantees, and contact-rich manipulation remains future work. Hardware trial counts, success rates, training cost and code availability: Not reported in the paper. Hardware π₀ inference used an external RTX 5090 workstation connected by wired Ethernet.

Industry impact

Developers of quadrupedal mobile manipulators and teams integrating VLA policies could use this interface for research prototyping now. Product use would require validation on the intended arm, base and operating conditions, with quantified hardware reliability and support for the required contact interactions. A commercial deployment timetable: Not reported in the paper.

The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.