💻 computer science

Calf-Integrated Arms for Bimanual Quadruped Loco-Manipulation

This paper presents a bimanual quadruped robot design that integrates arms into the front calves of a Unitree Go2, enabling stable, two-handed manipulation at ground level without rearing, while utilizing a vision-language model for long-horizon autonomous task execution.

Original authors: Yan Pan, Yuanchuan Ren, Chipui Chan, Jingcheng Sun, Chengxu Zhou

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yan Pan, Yuanchuan Ren, Chipui Chan, Jingcheng Sun, Chengxu Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a four-legged robot dog, like the popular Unitree Go2. Usually, if you want this robot to pick things up, engineers have to bolt a robotic arm onto its back (like a backpack). But that arm sits high up, making it hard to reach things on the floor without the robot awkwardly stretching down. Alternatively, some designs try to use the robot's own legs as hands, but that means the robot has to stand on its hind legs to use its front paws, leaving it unstable and unable to walk while working.

This paper introduces a clever new design: giving the robot "hands" built directly into its front shins (calves).

Here is how it works, broken down into simple concepts:

1. The "Shin-Arm" Transformation

Instead of a backpack arm, the researchers replaced the lower part of the robot's front legs with a special module. Think of it like a telescoping fishing rod that can also bend and twist.

The Slider: A part of the leg can slide up and down (like a telescope extending), allowing the "hand" to reach all the way to the floor.
The Joints: The arm can pitch (tilt forward/back) and yaw (twist left/right).
The Result: The robot can now grab things right off the ground while keeping all four feet firmly planted on the floor. It doesn't need to stand on its hind legs or stretch its neck awkwardly.

2. The "Two-Handed" Dance

Because both front legs have these built-in arms, the robot can use both hands at once.

The Overlap: The arms can twist inward so that the left and right "hands" meet in the middle of the robot's body.
The Magic: This allows the robot to do things like lift a heavy box with two hands, pass an object from one hand to the other, or carry a basket in one hand while opening a door with the other—all while walking normally with all four feet on the ground.

3. The "Brain" (The Vision-Language Model)

The robot doesn't just follow a pre-written script. It uses a "brain" powered by a Vision-Language Model (a type of AI that understands both images and text).

How it thinks: You give the robot a simple instruction like, "Put the basket in the cabinet."
The Process: The robot looks at the room through its camera. It breaks the big task into small steps (skills) it already knows, like "walk to cabinet," "open door," "grab basket," and "place basket."
Autonomy: It decides the order of these steps on its own. If the door is closed, it opens it. If the basket is far away, it walks to it first. It acts like a smart assistant that figures out the "how" based on what it sees.

4. What They Tested (In Simulation)

The researchers tested this design in a computer simulation (a video game world) because building a physical prototype takes time. They successfully demonstrated three complex tasks:

The Cabinet Task: The robot walked to a cabinet, carried a basket in one hand, opened the door with the other, and put the basket inside.
The Two-Handed Lift: Both arms grabbed a box and lifted it together.
The Handover: One arm picked up an object and passed it to the other arm without dropping it.

5. The Trade-Offs and Limits

The paper is honest about what this design can't do yet:

No Twisting: The gripper (the "fingers") can't rotate around its own axis. So, if you need to grab a kettle by a handle that requires twisting your wrist, this robot can't do it yet.
The "Marker" Problem: To grab things, the robot currently needs a bright red sticker (a marker) on the object to know exactly where to grab. It can't just look at a random object and guess the best grip point yet.
Simulation Only: All these results happened in a computer simulation. The robot hasn't been built and tested in the real world yet.

Summary

This paper proposes a robot dog that doesn't need a backpack arm or a "stand-up" pose to work. Instead, it has built-in, sliding arms in its front legs that let it grab things from the floor with two hands while walking. It uses an AI "brain" to figure out how to chain these actions together to complete complex chores, like tidying up a room. It's a step toward making quadruped robots true helpers for messy, human-scale tasks.

Technical Summary: Calf-Integrated Arms for Bimanual Quadruped Loco-Manipulation

Problem Statement
Current quadruped loco-manipulation designs face a fundamental trade-off between manipulation capability and stance stability. Trunk-mounted arms typically reach high, making ground-level grasping difficult and usually limiting the robot to a single manipulator. Utilizing legs as manipulators often requires lifting a foot from the ground, compromising the support polygon, or rearing onto hind legs (as in the LocoMan design), which fixes the base in place. Consequently, no existing design allows a quadruped to perform bimanual, ground-level manipulation while maintaining a stable, four-footed stance with a base free to walk. Furthermore, existing systems often lack high-level planners capable of autonomously sequencing long-horizon tasks based on visual and state feedback.

Methodology
The authors propose a morphological design and control framework that integrates manipulators directly into the front calves of a Unitree Go2 quadruped.

Mechanical Design: Each front leg is modified to include a seven-degree-of-freedom (DoF) system. The original three leg joints are retained for locomotion, while four new joints form the manipulator:
- $q_1$ (Prismatic Slider): A 0.105 m stroke slider housed inside the rebuilt calf, driven by a non-backdrivable stepper motor to lift objects from the ground.
- $q_2$ (Pitch): A revolute joint for fine approach control.
- $q_3$ (Yaw): A revolute joint rotating about the vertical axis, sweeping the gripper toward the body centerline to enable the left and right arms to overlap their workspaces.
- $q_4$ (Gripper): A parallel-jaw gripper with a 4 cm opening.
  This configuration allows the robot to grasp ground-level objects with both arms while all four feet remain planted.
Control Architecture: The system employs a two-layer control architecture:
- High-Level Planner: A Vision-Language Model (VLM), specifically Kimi K2.6 accessed via cloud API, selects discrete skills from a predefined library ( $S$ ). The planner conditions its selection on the head-mounted RGB-D camera image and a structured world state (tracking task progress flags). It queries the VLM only at skill boundaries, not at every control step.
- Low-Level Execution: Selected skills are executed via Finite State Machines (FSMs). The FSMs coordinate a learned locomotion policy (trained via Proximal Policy Optimization in Isaac Lab) for the 12 leg joints and a 3-DoF Damped Least-Squares (DLS) Inverse Kinematics (IK) solver for the arm joints ( $q_1$ – $q_3$ ). The gripper ( $q_4$ ) is commanded directly.
- Perception: Grasping relies on detecting a red marker on the object. A depth-based back-projection algorithm locates the 3D grasp point, filtering out partial detections near image borders.

Key Contributions
The paper makes two primary contributions:

Integrated Hardware: The design of a manipulator integrated into the front calf of a quadruped, enabling bimanual, ground-level grasping while maintaining a four-footed stance. This trades a smaller bimanual workspace (compared to rearing designs) for stability and a walk-ready base.
Autonomous Planning: The integration of a VLM to sequence skills from a predefined library based on visual input and task state, demonstrated on a long-horizon bimanual task in simulation.

Results
The design was evaluated entirely in simulation (Isaac Lab for training, MuJoCo for evaluation) across three bimanual tasks:

Long-Horizon Cabinet Task: The robot carried a basket in one arm, walked to a cabinet, opened the door with the other arm, and placed the basket inside. The VLM autonomously selected the skill sequence (grasp, walk, open, place) based on visual feedback.
Cooperative Two-Handed Lift: Both arms grasped a single box by its side handles and lifted it simultaneously, utilizing the workspace overlap provided by the yaw joints.
Inter-Arm Handover: The left arm transferred an object to the right arm at the body centerline without placing it on the ground.

In the cabinet task, the robot completed the sequence in 18.6 seconds of motion time. The system demonstrated robustness to depth sensor noise, maintaining a 70–74% grasp success rate with up to 10 mm standard deviation noise. The VLM queried only four times for the entire task, with latency occurring while the robot held a stable stance.

Significance and Limitations
The paper claims significance in demonstrating a quadruped capable of complex, bimanual, ground-level manipulation without sacrificing locomotion stability or requiring the base to be fixed. It highlights that the hardware design enables a planner to execute long-horizon tasks with all four feet planted.

The authors explicitly note several limitations:

Kinematic Constraints: The arm lacks a roll joint, preventing the gripper from rotating around its approach axis, which limits grasping objects with specific handle orientations (e.g., a kettle).
Skill Library Coverage: The VLM can only select from predefined skills; it cannot generate new behaviors or states. Tasks requiring re-grasping at new points, in-hand rotation, or dynamic obstacle avoidance are currently infeasible.
Perception Dependency: The system relies on fiducial markers for grasp points. Learned detectors (YOLO-World, SAM) currently fail to provide precise grasp coordinates for complex objects, necessitating a dedicated grasp-synthesis network for future work.
Simulation Only: All results are from simulation. The authors state that building a physical prototype to test sim-to-real transfer is the immediate next step.

Future work suggested by the authors includes physical prototyping, developing grasp-synthesis networks to replace markers, and exploring secondary use cases such as using the prismatic slider as a leg extension for stilt-like locomotion or lifting objects onto the robot's trunk.