TL;DR
This paper introduces a holistic "shared-waist" calibration method for wheeled humanoid robots that accounts for both geometric misalignments and joint flexibility (compliance). By optimizing the parameters of both arms and the shared waist simultaneously—and using a gravity model to estimate waist loads where sensors are missing—the researchers reduced end-effector errors from over 30 mm down to roughly 1 mm.
What Problem Does This Paper Address?
Wheeled humanoid robots typically feature two arms mounted on a single actuated waist. For these robots to perform precise dual-arm tasks, both hands must be accurate within a shared coordinate system. However, two major hurdles exist:
- Shared Error Sources: Any error in the waist model (like a slight tilt or joint flexibility) propagates to both arms. Traditional calibration often treats each arm as an independent chain, leading to two conflicting models for the same waist.
- Missing Sensors: While arm joints often have torque sensors to measure load-induced bending (compliance), waist joints usually do not. This makes it difficult to account for how the robot’s own weight and payloads cause the structure to sag or deflect.
The authors aim to bridge this gap by providing a unified calibration framework that treats the robot as a single kinematic tree rather than a collection of independent branches.
Methods
The authors propose a two-stage elasto-geometrical calibration procedure:
- Stage I (Extrinsics): The robot moves individual joints in arcs. An optical tracker measures these movements to identify the tracker-to-base transform and the target offsets on the robot's flanges using circle point analysis.
- Stage II (Whole-Body Identification): This is the core contribution. The system identifies a "complete" kinematic model (using Modified Denavit-Hartenberg parameters plus extra parameters for parallel axes and branches) and a compliance model.
- Shared-Waist Model: Instead of calibrating arms separately, they solve one large optimization problem where the waist parameters are shared. Mathematically, they show that the information gained about the waist is the sum of the information from both arms (via Schur complements).
- Gravity-Driven Compliance: Since the waist lacks torque sensors, the authors use a CAD-based mass model to calculate the gravity torques acting on each joint. They model both axial twisting and transversal bending (tilts).
- MAP Estimation: To prevent the model from "overfitting" to noise—especially when few calibration poses are available—they use Maximum A Posteriori (MAP) estimation with Gaussian priors. The scales of these priors are automatically tuned to maximize the evidence of the training data.
Datasets
The method was validated on two production robots from AgiBot:
- AGIBOT G2: A humanoid with a complex 5-DOF revolute waist.
- AGIBOT G2max: A humanoid with a simpler 1-DOF revolute pitch waist.
- Conditions: Both robots were tested in "no load" and "loaded" (5kg or 10kg per arm) configurations.
- Ground Truth: Measurements were taken using a high-precision optical 3-D tracker (FreeScan Trak Nova).
Key Insights and Ideas
- The Power of Duality: Calibrating both arms together isn't just a matter of convenience; it significantly reduces uncertainty. The two arms often contribute "complementary" information about the waist's orientation and position.
- Waist Compliance is King: The biggest accuracy gains came from modeling waist flexibility. Without it, geometric parameters try to "absorb" the bending error, leading to a model that only works in specific poses.
- Sequential vs. Simultaneous: The authors proved that "sequential" calibration (fixing the waist with one arm, then calibrating the second) loses accuracy for the second arm. Independent calibration is even worse, producing two different waist models that can't be used at the same time.
Key Results
- Massive Error Reduction: On the G2 robot, mean position errors dropped from 11.3–31.1 mm (nominal model) to 0.50–1.11 mm after calibration.
- Outperforming State-of-the-Art: The proposed method (DUAL-GC) showed 17–63% lower mean errors compared to adaptations of existing whole-body calibration methods (like those used for the TALOS biped).
- Efficiency: With just 20 training poses, the DUAL-GC method was significantly more stable than independent calibration, which saw errors spike due to the difficulty of identifying a 5-joint waist from a single chain.
Limitations and Future Work
- Mass Model Dependency: The waist compliance relies on a CAD-based mass model. While effective, errors in the CAD model (like unknown cable weights) could leak into the compliance parameters.
- Static vs. Dynamic: The calibration is performed in static poses. While this covers gravity-induced bending, it doesn't account for dynamic deflections during high-speed movement.
- Future Directions: The authors plan to automate the waist extrinsic identification (Stage I) for the G2max and test how well compliance parameters identified with one payload generalize to different loads.
Summary updated