TL;DR
This paper presents a force-controlled soft robotic hand that uses integrated bend and force sensors to achieve safe grasping of delicate objects. By developing a data-driven method to accurately estimate contact forces and employing a PI controller, the hand can grasp deformable and fragile items without causing damage or dropping them.
What Problem Does This Paper Address?
Grasping delicate or deformable objects is a significant challenge for robotic hands. While soft robotic hands offer inherent compliance, making them seem ideal for safety, they can still crush objects if the applied force is too high. Existing methods for force sensing in soft hands are often insufficient, especially for small objects, or require custom-fabricated sensors. This paper addresses the need for a reliable and accurate way to control grasping forces in soft robotic hands to ensure both stability and safety when interacting with a wide range of objects.
Methods
The core of this work lies in integrating off-the-shelf resistive bend and force sensors (FSRs) into a soft robotic hand's fingers. To overcome the inherent inaccuracies of FSRs when bent, the authors propose a data-driven calibration method. This method involves:
- Internal Force Characterization: Actuating the soft finger in free space and measuring the force output at various bending angles. This data is used to build a model of the "internal force" generated by the sensor itself due to bending, independent of external contact.
- Contact Force Estimation: The actual contact force ($F_c$) is estimated by subtracting the learned internal force ($F_i$) from the raw measured force ($F_m$) from the FSR: $F_c = F_m - F_i$.
- Force Control: A Proportional-Integral (PI) controller is used to regulate the pneumatic pressure applied to the actuators. The controller takes the error between the desired contact force and the estimated contact force as input and outputs a Pulse Width Modulation (PWM) signal to control the actuators. A derivative action was omitted due to its sensitivity to noisy sensor readings.
The hand design itself is based on the RBO Hand 2 for its kinematic structure, allowing for power grasping, but adopts a parallel finger arrangement similar to the DRL Soft Hand for easier sensor integration.
Datasets
The paper does not explicitly mention using pre-existing datasets for training or evaluation. Instead, it describes experimental setups where data was collected specifically for this research. This includes:
- Internal Force Data: Force measurements from each finger at various bending angles in free space, used to train the internal force model.
- Contact Force Estimation Data: Measurements of estimated contact force versus actual contact force at different finger configurations and distances.
- Force Controller Response Data: Time-series data of contact force response to step reference signals.
- Grasping Performance Data: Quantitative results (drop percentage, deformation percentage) for grasping various deformable and fragile objects under different force set-points.
Key Insights and Ideas
- FSRs can be made reliable with proper calibration: The paper demonstrates that even simple, off-the-shelf FSRs, which are notoriously noisy and inaccurate when bent, can be made to provide reliable force measurements through a data-driven internal force characterization and subtraction method.
- Internal force is a critical factor in soft sensor readings: The significant deviation of FSR readings from zero when bent in free space highlights the importance of accounting for this internal force for accurate contact force estimation.
- Simple PI control is sufficient for stable grasping: The use of a basic PI controller, rather than a more complex PID, proved effective in maintaining a constant grasping force, suggesting that for this application, the added complexity of a derivative term might not be necessary and could even be detrimental due to noise.
- Force control is crucial for deformable objects: Even with passive compliance, soft hands can easily deform or damage objects if the grasping force is not regulated. This work shows that active force control is essential for safe interaction.
- Object property estimation potential: The paper hints at the possibility of using the integrated sensors to estimate object properties like hardness by analyzing the relationship between pressing force and finger deformation.
Key Results
- Contact Force Estimation Accuracy: The estimated contact force varied only slightly (0.01 to 0.15 N) around the actual contact force, demonstrating sufficient accuracy for grasping.
- Force Controller Performance: The force controller closely followed step reference signals with a Root Mean Square (RMS) error of 0.1 N and a settling time of approximately 400 ms. The controller also successfully regulated pressure to achieve a target contact force of 2.5 N without overshoot.
- Safe Grasping: The force-controlled soft hand successfully grasped deformable objects (plastic cup, paper cup) and a fragile object (eggshell) without dropping, deforming, or breaking them across a range of tested force set-points. For instance, a plastic cup could be deformed with as low as 1 N of force, and the controller prevented this.
- Object Property Discrimination: The system could distinguish between a stiff object (spray can) and a soft object (woolly hat) based on the relationship between bending angle and estimated contact force.
Limitations and Future Work
Limitations:
- Sensor Noise: While the PI controller was designed to be robust to noise, fluctuations in estimated contact force still led to some duty cycle variations.
- Custom Fabrication: While off-the-shelf sensors were used, the integration and calibration process still requires careful implementation.
- Limited Object Set: The safe grasping experiments were conducted on a limited set of objects.
Future Work:
- Learning Dynamics Models: The authors suggest exploring the possibility of learning accurate dynamics models from the force and bend data.
- Dexterous Manipulation: Investigating how these models can be used for more dexterous manipulation tasks, potentially through model predictive control or model-based reinforcement learning.
- Human-Robot Interaction: Leveraging the safety features of soft hands and advanced control for more complex human-robot interaction scenarios.
- Object Property Estimation: Further developing the object property estimation capabilities to more accurately characterize hardness and other material properties.
Summary updated