Introduction to LingBot-Touch
On November 11, 2023, the Chinese company Robbyant launched LingBot-Touch, an open tactile perception model designed to enhance robotic task performance through sophisticated touch sensors. This cutting-edge system integrates multiple tactile sensing technologies, harmonizing data from diverse sensors to simplify system adaptation and improve reliability.
LingBot-Touch excels at detecting touch via various sensor types and demonstrated a 14 to 26 percentage point improvement in task success rates during trials. The model was trained using nearly 5,000 hours of real-world physical contact data and incorporates three tactile sensing methods: magnetic, visual, and piezoresistive—the latter identifying touch by monitoring changes in electrical resistance.
Key Benefits of LingBot-Touch
By converting signals from different sensors into a unified format, LingBot-Touch effectively overcomes previous issues related to incompatible data streams that hampered system retraining. It analyzes contact location, force distribution, and their temporal variations, enabling more nuanced tactile understanding. Training data can be combined from robotic hands and tactile gloves, further boosting operational efficiency.
During 26 real-world tests, two robotic hands—Sharpa Wave and XHand—were evaluated alongside four control models: LingBot-VLA, π0.5, Diffusion Policy, and TRex. Without tactile feedback, the Sharpa Wave with the TRex controller achieved a 56.25% success rate. Providing raw sensor signals improved this to 62.5%, while processing with LingBot-Touch raised accuracy dramatically to 90%.
The XHand robotic hand utilizes magnetic sensors absent from LingBot-Touch's original dataset. To accommodate this, training incorporated tactile glove data without altering LingBot-Touch’s parameters, modifying only the control model. There are two LingBot-Touch versions tailored for different computational capacities; all tests employed the lightweight Tiny variant, featuring around 5.3 million parameters. Integrating Tiny increases control model sizes by just 0.13–0.16%, while the larger version holds promise for enhanced performance with expanded training data.
LingBot-Touch by Robbyant marks a major advancement in tactile perception technology for robotics, promising significant improvements in robotic performance and adaptability across various applications.
The integration of tactile sensors into robotics stands to revolutionize automation across sectors from manufacturing to healthcare. LingBot-Touch’s ability to process and interpret touch information offers the potential to elevate robotic system efficiency, opening doors to applications in complex environments where precision and adaptability are essential.
As the field of robotics advances, innovations like NVIDIA's latest model further emphasize the importance of integrating human-like skills into robotic systems. Understanding how these developments complement technologies such as LingBot-Touch can provide deeper insights into the future of automation and tactile interaction.