David Buergel’s Post

TossingBot: Learning to Throw Arbitrary Objects Researchers have trained a robot to throw objects using visual observations and trial and error, overcoming challenges posed by varied object poses and properties. By combining simulation and deep learning, the system achieves accuracy and generalization for new objects and target locations. This technology might find applications in sorting rubbish or kitting non-sensitive parts, allowing robots to rapidly place objects beyond their kinematic range in pick-and-place scenarios. Thoughts ⁉️ #robotics #simulation #gripper #research #cobots Via: Andy Zeng from Princeton University Inspired by: Daniel Kuepper

Integrating the TossingBot technology into a business network https://t.me/theunityhub can provide several benefits: Efficiency: TossingBot enables robots to handle arbitrary objects with accuracy and generalization, streamlining sorting and kitting processes in various industries. By efficiently placing objects beyond their kinematic range, robots can optimize pick-and-place scenarios, enhancing overall operational efficiency. Flexibility: The ability of TossingBot to learn from visual observations and trial and error allows robots to adapt to diverse object poses and properties. This flexibility enables robots to handle a wide range of objects without the need for extensive reprogramming, facilitating flexible and adaptable automation solutions.

the RobotGPT is moving from science fiction to reality! 🚀 The fusion of perception, planning and control with foundation models demonstrates a promising approach but it requires grounding policies. 🤖 Combining different data sources, such as simulation data, real-world data, and language models is crucial for behaviour cloning in robotics, as it provides the necessary training data for robots to learn complex tasks. 🗣 Shoutout to Google DeepMind, NVIDIA, AI at Meta, and Physical Intelligence for their contributions + Covariant, Everyday Robots, and Skild AI for their bold vision in building general-purpose robots. 💯🎯 #Robotics #Manipulation #Navigation #AI #MachineLearning #Data #FoundationModel #VLM #VLA

In the video, two robot arms are fully autonomous and controlled by a single neural network conditioned on different language instructions (four axes and five axes robot arms). The other two arms are teleoperated. The entire hardware cost in the video is about $1400. The video is at 1.5x speed. Video Source: Tau Robotics / Alexander Koch . . . #Robots #Engineering #Manufacturing #Research #Automation #AI #ArtificialIntelligence #Technology #Tech #Robotics #innovation #development 

𝐀𝐝𝐯𝐚𝐧𝐜𝐢𝐧𝐠 𝐂𝐡𝐚𝐫𝐚𝐜𝐭𝐞𝐫 𝐂𝐨𝐧𝐭𝐫𝐨𝐥 𝐟𝐨𝐫 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝐑𝐨𝐛𝐨𝐭𝐬 Recent breakthroughs in physics-based character control are helping us learn from unstructured motion data. However, it’s still a challenge to create a single control policy that works with diverse motions and can be used on real-world robots. In a new approach, researchers developed a two-stage method to address this: Stage 1: Train a variational autoencoder (VAE) to extract a latent space encoding from unstructured motion data, capturing complex movement patterns. Stage 2: Use the learned latent code to train a policy that maps kinematic input to dynamics-aware output, enabling accurate motion imitation. This separation allows the model to learn better, avoid errors, and adapt to unseen motions. The method has been tested in simulations and on bipedal robots, successfully bringing complex, dynamic movements into the real world. Please check reference for my post in comment section #Robotics #AI #PhysicsBasedControl #MachineLearning #RobotMotion #ImitationLearning #CharacterControl

UC Berkeley Pushes Boundaries of #Robotic Dexterity Researchers at the University of California, Berkeley have made significant strides in enabling #robots to mimic human movements with remarkable precision. Overcoming hardware limitations, they developed #HATO, an affordable #teleoperation system for two-arm setups with multi-fingered hands. Adapting #prosthetic hands with touch sensors, the team harnessed visuotactile data to train #AI models on intricate, dexterous tasks. Their comprehensive analysis highlights the impact of dataset size, sensing modalities, and visual processing on policy learning. This groundbreaking work paves the way for advanced #robotic manipulation capabilities, unlocking new frontiers in real-world applications. #robotics #ai #artificialintelligence #winepouring #QuesationForGroup

Immerse yourself in the world of Humanoid Teleoperation for real-world data collection! 🤖🌐 Explore our latest work in advancing robot technology through iterative learning of control algorithms. Check out the details here: [Humanoid Teleoperation to Collect Data for Learning](https://lnkd.in/gvq8EPa7). 🚀💡 Join us in propelling the progress of General-Purpose Humanoid Robots, pushing the boundaries of teleoperation, and enriching training data for continuous algorithmic refinement. #GeneralPurpose #HumanoidRobot #Teleoperation #TrainingData #Algorithm #LimXDynamics #Innovation More Contents: Woongsik Dr. Su, MBA

This year at NeurIPS, we gave an Oral presentation on a new AI diffusion method for automatically designing soft robot bodies and behaviors. DiffuseBot uses the same technique that powers image generators like Dall-E, but generates soft robots rather than images. We co-optimize shape, actuator placement, and control to solve tasks in passive dynamics, manipulation, and locomotion. One framework does it all. Because we insert differentiable physics directly into the computational pipeline, we don't need big datasets of robot designs to make this work --- which is great, because those datasets don't exist! DiffuseBot improves from its own trial and error, and because it's constantly testing what it generates, you don't have to worry about weird "hallucinations" or other common artifacts of generative AI. Watch the CSAIL video on the research: https://lnkd.in/edhpqkkG Or check out more details here: https://lnkd.in/erz2qPw7 Joint work with Tsun-Hsuan Wang, Juntian Zheng, Pingchuan Ma, Yilun Du, Byungchul Kim, Joshua B. Tenenbuam, Chuang Gan, and Daniela Rus

In the world of robotics, teleoperation plays a crucial role in enhancing robot learning, especially through imitation learning and human-in-the-loop corrections. In our recent work, “VITAL: Visual Teleoperation to Enhance Robot Learning through Human-in-the-Loop Corrections,” we explore how low-cost visual teleoperation can be harnessed for imitation learning. This innovative method bridges the gap between real and simulated environments, accelerating learning and making robots more capable and versatile. It also enables real-time human corrections, boosting adaptability and efficiency in diverse tasks. Check out the video of our experiments and the paper to know more about this work. - Preprint: https://lnkd.in/eMCfxz7D - Video: https://lnkd.in/ewW6bAgw #Robotics #AI #MachineLearning #Teleoperation