Quadrupedal mobile manipulators can extend their manipulation workspace by coordinating the mounted arm with the mobile base. At the task level, short-horizon end-effector (EE) trajectories provide an interface for specifying the desired manipulation motion without prescribing how it should be realized by the arm and base. However, executing such trajectories remains challenging because the controller must determine when and how the base should move, and this decision depends on the mounted-arm morphology. We therefore present PolyMorphWBC, a hierarchical controller that takes an EE trajectory as the sole task-level motion command and generates coordinated arm--base motions across four families of 6-DoF mounted arms, with extended-reach trajectories beyond the initial fixed-base reach. Within PolyMorphWBC, a trajectory-conditioned coordinator determines the base motion from previewed EE targets, while arm-specific Inverse Kinematics (IK) tracks the current target, and a frozen locomotion backend executes the resulting motion. Cross-arm reuse is enabled by a Unified Robot Description Format (URDF)-derived morphology conditioning, together with online residual-wrench feedback. We train PolyMorphWBC on procedurally generated arms and structured trajectories within and beyond fixed-base reach. In IsaacLab, it achieves 93.95% completion with 2.30cm position error and 4.81° orientation error on held-out trajectories and arm morphologies. In MuJoCo evaluation against UMI-on-Legs, PolyMorphWBC reduces EE position error by 69--72% and orientation error by 91--97% across four within- and extended-reach trajectories. On a physical Go2--ARX5 system, we demonstrate language-conditioned extended-reach pick-and-place using EE trajectories from pi_0 without vision-language-action (VLA)-generated base commands.
Figures & tables
Fig. 1: Cross-arm reuse and extended reach with PolyMorphWBC. Top: a shared control hierarchy maps short-horizon EE trajectories to morphology-dependent arm–base motion across mounted-arm morphologies. (a) Representative arms exhibit different fixed-base workspaces. (b) Coordinated base motion enables tracking of EE trajectories extending beyond the arm’s initial fixed-base reach, without task-level base-motion commands. Panel (b) is schematic.
Fig. 2: Pipeline of PolyMorphWBC. (a) The coordinator converts an EE trajectory preview into a base command, while arm-specific IK tracks the current target and the locomotion backend produces leg actions. (b) The morphology encoder is pretrained with kinematic and residual-wrench supervision and then frozen. (c) The backend and its morphology projection are trained with PPO under sampled base commands and independently executed, offline-IK-based arm motion, using target–wrench–response history. (d) With the encoder and backend frozen, the coordinator and its projection are trained through sampled EE trajectory covering motion within and beyond initial fixed-base reach. Dashed arrows indicate optimization updates.
Code
Kinematic family
Anchor arm
PA
Three parallel axes
ARX X5
SW
Spherical wrist
PiPER
OW
Offset wrist
Compact xArm6
CW
Curved wrist
Compact MICO
TABLE I: Mounted-arm kinematic families and their anchor models.
All trajectories
Extended reach ( x>1m )
Extended reach ( x<0m )
Family
Rcomp (%) ↑
ep (cm) ↓
eR ( ∘ ) ↓
Rcomp (%) ↑
ep (cm) ↓
eR ( ∘ ) ↓
Rcomp (%) ↑
ep (cm) ↓
eR ( ∘ ) ↓
PA
98.45
1.62
3.06
97.57
1.35
2.43
99.04
1.80
3.12
SW
98.65
1.96
4.10
99.65
1.58
3.05
97.81
2.03
4.09
OW
91.03
2.75
6.40
94.43
2.12
4.81
88.39
2.60
5.29
CW
87.65
2.88
5.68
87.65
2.70
5.15
81.01
3.20
6.23
Overall
93.95
2.30
4.81
94.83
1.94
3.86
91.56
2.41
4.68
TABLE II: EE-trajectory tracking on held-out trajectories and arm morphologies. The two subsets contain trajectories with at least one reference target satisfying the indicated x condition in the initial base frame.
Anchor arm (family)
Rcomp (%) ↑
ep (cm) ↓
eR ( ∘ ) ↓
ARX X5 (PA)
99.30
1.57
2.96
PiPER (SW)
93.16
1.98
4.45
Compact xArm6 (OW)
99.27
1.93
4.09
Compact MICO (CW)
86.62
2.67
5.26
TABLE III: EE-trajectory tracking on anchor arm models without policy fine-tuning.
PolyMorphWBC
UMI-on-Legs [ 11 ]
Trajectory shape
ep (cm) ↓
eR ( ∘ ) ↓
ep (cm) ↓
eR ( ∘ ) ↓
Vertical circle
1.38
2.06
5.00
22.72
Vertical square
1.43
2.14
5.17
23.76
Horizontal rectangle
1.48
2.01
4.77
40.41
Horizontal base orbit
2.04
2.38
7.31
76.57
TABLE IV: Comparison of EE-trajectory tracking errors on Go2–ARX5 in MuJoCo.
Fig. 3: EE-trajectory tracking on Go2–ARX5 in MuJoCo. Panels (a)–(d) show the vertical circle, vertical square, horizontal rectangle, and horizontal base orbit, respectively. Black dashed curves show commanded trajectories; blue and orange solid curves show trajectories executed by PolyMorphWBC and UMI-on-Legs, respectively. Color-matched arrows show EE local z -axes projected onto each plotting plane.
Fig. 4: Real-world demonstrations of PolyMorphWBC on a Go2–ARX5 platform. (a) EE-commanded teleoperation, where straight red arrows indicate the commanded EE motion and curved red marks indicate the controller-generated base motion. (b) VLA-driven extended-reach pick-and-place using π0 . Each row shows successive stages from left to right. In both cases, only the EE motion is commanded at the task level; no base-motion command is provided by either the operator or π0 .
ΔRcomp (pp)
Family
All trajectories
x>1m
x<0m
PA
−0.48
−1.22
−0.14
SW
+0.55
+0.52
+0.41
OW
+2.59
+2.26
+3.18
CW
+9.90
+9.52
+8.24
TABLE V: Ablation on morphology conditioning on held-out arms. ΔRcomp (pp) denotes the completion rate of the full model minus that of the variant without morphology conditioning.
Shanghai Research Institute for Intelligent Autonomous Systems Shanghai 201210, China · College of Electronics and Information Engineering (Tongji University) Shanghai 201210, China
Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA-15213 USA · Mitsubishi Electric Research Labs (MERL), Cambridge, MA-02139 USA