Visible-Reachable Workspace for Perception-Aware Humanoid Design

A humanoid can usually reach far more space than it can see. The visible-reachable workspace (VRW) measures the part it can both reach and see, at design time. Applied to this robot, it led to two independently aimed camera modules instead of one shared neck.

Abstract

Workspace analysis measures where a robot can place its end effector. For visually guided manipulation, reachability alone is insufficient: a kinematically reachable target may not be visible in the specific pose required to reach it. The robot must then redirect its sensing or move its body to acquire a view, turning a perception limitation into additional motion. Existing humanoids largely inherit this limitation when copying human form factors. We introduce the visible-reachable workspace (VRW), a design-stage measure that conditions visibility on feasible reaching configurations and extends it to concurrent visibility of spatially separated work regions. We apply VRW by building a 31-DoF humanoid with independently actuated RGB-D cameras. On the same robot, camera articulation increases visible-reachable coverage from 38% to 97%. With actuated camera layouts, a second camera raises pairwise coverage from 0.45 to 0.95, while a third changes it only to 0.97. In a controlled two-target reach-and-grasp benchmark, our dual-actuated design reduces mean completion time by 17% and mechanical energy by 19% relative to the same robot with its cameras fixed. Hardware experiments demonstrate simultaneous observation and manipulation of front/back and left/right target pairs without torso reorientation. The results suggest that reachability becomes a more informative design quantity for perception-driven humanoid manipulation when it is evaluated together with the sensing configurations that make the reachable space observable. We will open-source all software and the humanoid hardware design.

The Platform

VRW applies to any embodiment, and this robot is the platform built to study it. Two yaw-pitch camera modules sit on top of the torso, so gaze moves independently of torso heading. The shoulder pitch motor sits inside the torso to keep the shoulders narrow, and the shoulder and elbow ranges allow reaches in front of and behind the body. A whole-body policy sends residual joint-position commands to the legs, arms and cameras.

3×

The Measure

Reachability asks where the arm can put its hand. VRW keeps only the reachable points that a camera can also see from a configuration that achieves the reach. Joints needed for the reach are separated from joints still free to aim a camera, and manipulation takes priority. A wrist camera is modeled too, but its pose is tied to the reaching arm, so it adds no independent gaze. Whatever is left is reachable but blind.

Visible-reachable coverage, η = μ(W_VR) / μ(W_R), is the share of reachable volume a camera can see from a configuration that achieves the reach. Pairwise coverage, η₂, measures whether the robot can manipulate at one point while a camera still holds a view of another.

Six Embodiments

Under one common geometric evaluation, every external platform shows a gap between what the arms can reach and what the cameras can see. An actuated neck closes more of it than a fixed head. Cameras that share a neck still look in one direction, so concurrent views of two separated regions stay limited.

External models come from public simulation repositories. None of the external platforms can aim two camera views independently. Different embodiments induce different evaluation volumes, so these values are context and not a ranking.

Two-Target Reach and Grasp

The benchmark ran in physics simulation with the same whole-body policy and the same target-discovery planner for every camera configuration. Success rates were similar across all five configurations, but the time and energy needed were not. The actuated configurations spend much less time searching, because the robot finds a target by moving its cameras instead of walking until the target comes into view.

Act₂ against Fix₂: time −17%, energy −19%, search time −50%. Act₁ against Fix₁: time −24%, energy −35%, search time −82%. Values are success-conditional means over six scenarios, 30 trials per scenario (ten seeded layouts, three replicates each). Lower is better. Success rates ranged from 0.967 to 0.994.

All six scenarios under the four camera configurations. Shaded cones are the camera fields of view and green cubes are the targets.

Single fixedSingle actuatedDual fixedDual actuated

Single fixedSingle actuatedDual fixedDual actuated

Single fixedSingle actuatedDual fixedDual actuated

Single fixedSingle actuatedDual fixedDual actuated

Single fixedSingle actuatedDual fixedDual actuated

Single fixedSingle actuatedDual fixedDual actuated

On Hardware

We deployed the adopted configuration on the physical robot. With targets in reach, the two cameras observed a front/back pair and a left/right pair, and both arms grasped them without turning the torso. With targets out of reach, the robot located both first, then walked closer and grasped them. Each clip plays four trials together.

Continuous Operation

2×

Video

Interview

RealSense interviewed us about Duke Humanoid v2: watch on YouTube.

Links

• Project Website
• Paper (arXiv)
• PDF
• Code
• Open-source hardware

Citation

1
2
3
4
5
6
7
8
9

@misc{xia2026visiblereachableworkspaceperception,
      title={Visible-Reachable Workspace for Perception-Aware Humanoid Design},
      author={Boxi Xia and Zijiang Yang and Ryan Shin and Bokuan Li and Eric Wun-Hao Lu and Jacob Lee and Jiaxun Liu and Boyuan Chen},
      year={2026},
      eprint={2609.08905},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.08905},
}