🚀 Mujoco MJX now has Sphere-Cylinder collision working in the main branch! Just built it and ran MyoFinger & MyoHand on MPS via JAX (single sim is always slow). Looks like the next version will make MyoSuite fully compatible with MJX! 🔥🦾https://github.com/micropilot/myosuite-mjx (WIP) #MuJoCo #JAX
Future of Teleoperation & Prosthetics
The fusion of sEMG-based control + tendon-driven anthropomorphic robots opens doors to ultra-intuitive robotic hands for surgery, VR/AR, assistive tech, & space applications. This is the future! 🚀 #Robotics #NeuroTech
Seamless Mapping from Humans to Robots
By leveraging sEMG signals + tendon-based robotic actuation, we can create 1:1 mappings between human muscle activations and robotic movements, leading to naturalistic & highly responsive teleoperation. #MyoSuite
sEMG: The Key to Intuitive Control
Surface electromyography (sEMG) sensors, like Meta CTRL-Labs' wristband, capture muscle signals directly from the forearm, providing real-time, non-invasive control of robotic hands—eliminating occlusion issues in vision-based tracking. #emg2pose
shorturl.at/PUIJp
Why Tendon-Driven Hands?
Unlike rigid joint-actuated hands, tendon-driven robotic hands mimic human biomechanics, allowing for greater flexibility, force distribution, & adaptability in real-world tasks. Perfect for teleoperation and prosthetics! 🤖🖐️
The Human Hand: A Benchmark for Robotics
The human hand is the gold standard for dexterity & compliance. Its tendon-driven actuation enables natural, adaptive, and precise control something traditional robotic hands struggle to replicate. #MyoSuite #Robotics myosuite.readthedocs.io/en/latest/
We've expanded the emg2pose dataset to include 39-channel muscle-tendon control signals for @MyoSuite MyoHand model! 🚀 #AI #Robotics #RoboSoft25
I'm personally interested in how insights from the muscular system can inspire energy-efficient control for electric motors. This is especially relevant for today's humanoid robots, where battery life & control efficiency are critical. #ReinforcementLearning #Robotics
Why study muscle-based control in the first place? 🤔
Human muscles are energy-efficient & optimized for control precision. By learning from nature, we can design better, more efficient control policies for modern robots (e.g., humanoids, prosthetics).
In my experiments, I tested 8 RL methods (e.g., PPO, SAC, TD3) to train policies for accurate finger pose estimation. 🖐️
🎯 The goal: Move the finger into target poses using efficient muscle activations.
Spoiler: SAC performed best, closely followed by PPO!
Here’s a simple sketch showing the anatomy:
👉 Finger joints: MCP, PIP, DIP
👉 Tendons: flexion, abduction, extension
Each muscle-tendon unit works in opposition (like real muscles), providing a rich control challenge.
This intuitive design is based on Xu et al. (2012) [https://ieeexplore.ieee.org/document/6290710] and includes both biomechanical & robotic variants.
The MyoFinger model is a simplified representation of a human finger. It has 4 Degrees of Freedom (DoFs) and is controlled by 5 antagonistic muscle-tendon units. myosuite.readthedocs.io/en/latest/su...
