Here is the key idea of the video in a single sentence: Humanoid robots are rapidly advancing in design, capabilities, and functionality, but despite their impressive developments, they still face significant challenges and limitations that hinder their practical application and widespread adoption
Questions to inspire discussion
Manis Glove Technology
🖐️ Q: How does the Manis glove achieve accurate hand tracking?
A: The glove tracks 25 degrees of freedom using inverse kinematics based on 6DOF per fingertip (position and orientation), enabling accurate motion capture even when fingertips are obscured.
🔌 Q: What hardware enables the Manis glove's position tracking?
A: The system uses transmitters at the base and receivers in fingertips to determine precise fingertip position relative to the transmitter, with simple calibration allowing different hand sizes as long as sensors stay in place.
📳 Q: How does the Manis glove provide haptic feedback?
A: Haptic feedback at the PIP joints vibrates upon contact, enabling virtual world interaction and realistic surface contact simulation for teleoperation and clinical evaluations.
🎬 Q: What are the primary applications for the Manis glove?
A: After 10 years of development with Hollywood motion capture origins, the glove shows potential for high-quality clinical evaluation data, robotics, and medical fields.
Boston Dynamics Atlas Design
⚙️ Q: What actuator strategy does the new Atlas robot use?
A: Atlas features 100-degree rotary actuators for elbow and knee, using only two unique actuators to enable part reuse for left and right limbs, prioritizing cost, complexity, and reliability for factory use.
🔧 Q: How quickly can Atlas limbs be serviced?
A: Customers can swap limbs in under 5 minutes, with dramatically more reliable rotary actuators compared to linear ones, designed specifically for factory floors.
📦 Q: What are the trade-offs of reducing actuator SKU count?
A: Reducing SKU count lowers costs at large volumes but may lead to over or under-actuation in certain joints, impacting performance and load-bearing capacity.
🏭 Q: Why might Hyundai succeed at varied actuator production?
A: Hyundai's expertise in building lines and working with suppliers could enable production of varied actuator types at scale, balancing manufacturing trade-offs with performance needs.
Mechanical Design Considerations
🔄 Q: What are the issues with parallel mechanisms in humanoid waists?
A: Parallel mechanisms in humanoid waist joints are prone to failure and torque transparency issues, making direct drive systems preferable for better control and reliability.
💪 Q: How does the Cyber Robo humanoid's hip joint work?
A: The SPU drive in hip area uses parallel actuation mechanism for rotary abduction, flexion, and waist rotation, with two rotary actuators at top of hip joint visible in X-ray view.
🦵 Q: What mechanism does Cyber Robo use for knee actuation?
A: The knee actuator employs modified inverted Herkin's mechanism to linearize torque curve, with abductor actuator on inner thigh above knee and flexor actuator on back of thigh.
🤚 Q: How is the Cyber Robo hand actuated?
A: Forearm actuation utilizes 25 tendons actuated by single cylindrical unit, with five tendons per digit (two for abduction, three for flexion-extension in four fingers).
✋ Q: Why is wrist movement critical for humanoid dexterity?
A: Wrist movement is essential for dexterity in tendon-driven hands, requiring actuators in wrist and forearm for proper function, with wrist acting as central link in kinematic chain, combining flexion-extension and radial-ulnar deviation.
Robot Training and AI
🤖 Q: When is reinforcement learning necessary for humanoid training?
A: Reinforcement learning and simulation are necessary for training robots to perform non-human-like motions, as human data is insufficient for training beyond human capabilities.
🎯 Q: What does Boston Dynamics' new humanoid demonstrate about motion?
A: The robot demonstrates advanced symmetry, balance, and whole-body coordination, suggesting a shift towards robots inventing their own motion primitives unconstrained by human limitations.
🎮 Q: What are the limitations of simulation platforms for robot training?
A: Platforms like Aubot's Genie Sim 3.0 with NVIDIA Isaac SIM provide 10,000+ hours of synthetic data, but high-fidelity physics is still needed for accurate tactile feedback and locomotion.
⚠️ Q: Why do LLMs struggle with robot training scenarios?
A: Edge cases occurring once in a million are critical for training, but LLMs struggle to generate these rare unpredictable scenarios, limiting their effectiveness in robot training.
Dexterity and Sensing
🎯 Q: Can simple grippers match complex robotic hands?
A: AI-driven dexterous hands can outperform highly capable robotic hands through extensive training, as demonstrated by Aloha project where simple paddle grippers performed complex tasks like tying knots.
👁️ Q: When is proprioception more effective than vision?
A: Proprioception and touch sensing can be more effective than vision, as shown by DLR study where robotic arm with only joint torque sensors identified where a person was pushing, enabling tasks like writing.
🔍 Q: Why is sensor fusion crucial for robotics?
A: Sensor fusion is crucial as demonstrated by personal experience where closing eyes improved performance due to misleading visual input, highlighting importance of combining different sensory inputs.
🤲 Q: What sensing layers do robotic hands need?
A: Multi-layered sensing combining touch, proprioception, and joint sensing is necessary for tasks like feeling vibrations and detecting contact with threads, as vision alone has limitations.
🔩 Q: Is back-drivability a hardware or software property?
A: Back-drivability is primarily a software property, not hardware, as demonstrated by worm gear drive in Shankai hand, which is very hard to back-drive mechanically.
Robot Safety
⚠️ Q: What is the primary safety concern with robot motion design?
A: Safety is a major concern when allowing robots to design their own motions or forms, as there are countless potentially unsafe combinations that need extensive training to address.
🚫 Q: Should you catch a falling humanoid robot?
A: Never catch a falling robot; design robots to learn to fall safely and avoid catching them, as they can weigh 70kg and may flail unexpectedly due to software glitches.
📏 Q: What safe distance should robots maintain from humans?
A: Robots should maintain a 2-meter safe distance from humans; if a person encroaches, the robot must move away or emit a loud warning signal to avoid collisions.
🔊 Q: How should robots warn humans of their approach?
A: Robots should emit loud, attention-grabbing warning signals when approaching humans in urban environments, rather than annoying backup beeping used in cars.
🏭 Q: What should be prioritized over viral robot demos?
A: Robots must be designed with safety and reliability as top priorities, focusing on serviceability over viral moments, to prevent accidents and ensure safe operation in real-world environments.
Surgical Robotics
🏥 Q: When will surgical bots outperform human surgeons?
A: Surgical bots could outperform human surgeons in 10-20 years, but teleoperation and human oversight will remain crucial, as no fully automated surgical procedures exist today.
⏳ Q: What will delay widespread adoption of surgical bots?
A: Liability concerns, training requirements, and need for human-robot collaboration, especially with nervous child patients, will delay widespread adoption for at least 10 years.
🤖 Q: Will humanoid robots be optimal for surgery?
A: 3D printer-like automation with fixed tools may be more suitable for certain surgeries than humanoid robots, despite perception that humanoids are most capable robots.
👨⚕️ Q: Why is human presence still critical in surgery?
A: Human connection and trust in doctors are vital for patient cooperation, especially with children, and may be lost with fully automated surgical bots.
🖐️ Q: Can current humanoid hands perform surgery?
A: Surgical bots will likely be teleoperated by humans for a long time, as no humanoid robot hand has yet demonstrated the dexterity and precision required for surgery.
💰 Q: How will advanced surgical bots be distributed?
A: The first Optimus surgical bots that could outperform human surgeons will likely be limited in number (10,000) and expensive, raising questions about access and distribution of best surgical care.
Key Insights
Manufacturing and Design Trade-offs
🔧 Boston Dynamics' new Atlas robot uses only 2 unique actuators to minimize cost and complexity, with reliable rotary actuators replacing linear ones and enabling quick limb swapping in under 5 minutes for factory floor use
⚙️ Reducing SKU count in actuators for mass production can lead to overactuation or underactuation in humanoid robots, as seen with a large actuator at the ankle needed for ankle roll motion but creating suboptimal reflected inertia
🏭 Manufacturing scale impacts actuator design choices, with millions per year allowing more actuator types while thousands per year favors fewer types, leveraging Hyundai's expertise in building manufacturing lines
🔄 Boston Dynamics' new Atlas incorporates manufacturability and serviceability considerations with common parts for both left and right limbs, enabling cost-effective production and simplified maintenance
🦿 The new Atlas features a changed waist design and kinematics with individual actuators at knee and ankle, allowing for part reuse and cost reduction in manufacturing
Mechanical Limitations and Failure Points
⚠️ Parallel mechanisms like RSU at the waist can fail, lose torque transparency, and confuse actuators due to torque splitting between them, with linkages at the waist being a known weak point prone to failure
✅ Direct drive actuators are preferred for torque transparency and avoiding confusion in humanoid robots, eliminating torque splitting that can lead to confusion and balance loss when unexpected torque signals are encountered
🔩 The humanoid robot's hip joint uses parallel actuators for rotary abduction and flexion with separate waist rotation actuator, while the knee actuator employs a modified inverted Herkin's mechanism, a four-bar linkage, to linearize torque curve
🤔 The robot's forearm has 25 tendons actuated by a single cylindrical drive, creating a dead zone where tendons run freely, with unclear wrist joint mechanism raising questions about claimed 27 degrees of freedom
Hand Tracking and Dexterity
🧤 Manis glove uses inverse kinematics and 25 degrees of freedom to track finger position with 6DOF per fingertip using transmitters and receivers, achieving high precision even when fingers are obscured
📍 The Manis glove achieves accurate finger tracking without cameras or external sensors, using a kinematic model and inverse kinematics based on transmitter's position relative to fingertip receivers
🏥 Manis glove enables clinical evaluations to obtain high-quality data on patient hand function, with noted accuracy and ease of use for capturing hand gestures
🎴 The Wuji hand demo at CES used only two fingers (thumb and index) for all motions except stabilizing, showcasing impressive tactile sensors and dexterity through AI training for tasks like playing blackjack
🤖 AI-driven dexterity enables robotic hands with simple end effectors, like the Aloha project using paddle grippers, to perform complex tasks such as tying knots, demonstrating dexterity is more about AI training than physical capabilities
🦾 Proprioception and touch are crucial for robotic dexterity, as shown by DLR's robotic arm with force-torque sensors in joints determining where a person was pushing, even without touch sensing
👁️ Vision alone is insufficient for many robotic tasks, as demonstrated by need for multiple sensors when robot's hands obscure the object being manipulated
🔧 Back drivability in robotic actuators is primarily a software-based property, not hardware, as demonstrated by the worm gear drive in the Wuji hand making it very hard to back drive
Training and Simulation Challenges
🎮 Reinforcement learning and simulation are needed for training humanoid robots to perform non-human-like motions, as going beyond human capabilities lacks real-world training data, but the field still relies on motion capture and imitation learning
🔒 Safety is a concern when allowing robots to invent new motions, as there could be billions of unsafe combinations, making it easier to manually teach safe behaviors than training on all possible scenarios
🌐 LLMs can generate diverse training scenarios for robots but struggle with edge cases that happen rarely and aren't documented, requiring high-fidelity physics in simulators for realistic tactile feedback and locomotion
💻 Simulation platforms like Alibaba's Genie Sim 3.0, integrated with NVIDIA Isaac SIM, provide 10,000+ hours of synthetic data, 3D environments, and 2000+ tasks for comprehensive robot training, but still require significant post-training refinement
Safety Concerns
⚠️ Catching a falling robot is dangerous as 70kg metal can flail unpredictably due to software glitches while attempting to self-correct balance, making it as risky as catching a falling knife
🚨 Robots must learn to avoid humans within a 2-meter safe distance; if a person encroaches, the robot should move away or emit a loud warning signal, not just an annoying backup beep
👁️ Vision-based systems capable of distinguishing animate from inanimate objects will soon be cheap and widely available, but current standards require LiDAR, which is limited in object recognition
🎭 Spectacular robot demos at CES can obscure mechanical, safety, and simulation limitations; a robot's software may cause it to flail or self-correct in unexpected ways, posing safety risks
Medical and Surgical Applications
🏥 Surgical bots could outperform human surgeons in 10-20 years, but teleoperation and human oversight will remain crucial for safety, reliability, and serviceability, as no fully automated surgeries are currently performed without a human in the loop
⚖️ Liability concerns, training requirements, and the need for human connection in patient care will delay adoption of surgical bots, even if technically superior, with a potential 10-year lag before seeing them in practice
🤖 Humanoid robots are not the best form factor for surgery; 3D printer-like automation with specialized tools is more suitable, despite the psychological appeal of humanoid appearance to patients
🎮 Robotic surgery today is teleoperated by humans, not fully automated, with the surgeon controlling robotic arms, similar to laparoscopy with extensions for camera and tools
🧠 AI and LLMs could surpass human capabilities in decision-making, diagnostics, and imaging within a few years, but human oversight will still be needed for the foreseeable future
👶 Social interaction and human connection in patient care, especially with children, cannot be replaced by robots, as they have evolved over millions of years and are crucial for motivation and collaboration in treatment
Anatomical Design Details
🖐️ The humanoid robot's hand features five tendons per digit, with two for abduction and three for flexion-extension in the fingers, though the thumb may have a different configuration
🤸 The robot's wrist motion combines flexion-extension and radial-ulnar deviation, involving both the radiocarpal and midcarpal joints, unlike other humanoid robots that primarily focus on gripping and finger use
🦵 The humanoid robot's abductor is a dedicated actuator on the inner thigh, terminating above the knee, while flexion occurs on the backside, using a serial link mechanism differing from other robots' parallel mechanism
Market Positioning and Competition
📊 The Shankai Matrix robot, with a combination of linear and rotary actuators, showcased applications similar to other humanoid robots, but striking parallels in marketing material to Figure 03 and Optimus suggest intentional copying
🔍 The humanoid robot's actuators appear miniaturized and torque-dense, but it is unclear if the thin actuators can generate sufficient power for the legs and forearm, as other humanoid robots use larger actuators for this purpose
#Cobots #HumanoidRobots #OverTheHorizon
XMentions: @HabitatsDigital @overthehorizon @RoydenDesouza @GoingBallistic5 @anatomyumea @FarimaniMehrdad
WatchUrl: https://www.youtube.com/watch?v=CZoQ6jFR71w
Clips
-
00:00 🤖 Humanoid robot developers, like Boston Dynamics and Manis, are improving designs, scaling production, and enhancing tracking capabilities, but the technology is still evolving with trade-offs between performance, cost, and simplicity.
- Manis has developed a data glove that accurately captures hand gestures using a kinematic model and inverse kinematics, with potential applications in motion capture, robotics, and clinical evaluations.
- A haptic glove with embedded receivers, transmitters, and strain gauges provides precise hand tracking and realistic feedback, allowing users to feel virtual surfaces.
- The conversation clarifies the confusing naming situation between Meta's recent $2 billion purchase of Chinese AI startup Manis.ai and a Dutch company called Manis, which has a website at mananish meta.com.
- Humanoid robot glove developers are improving designs, scaling production, and working with third parties to enhance tracking capabilities, such as wrist position, without relying on inertial measurement units (IMUs) due to their tendency to drift over time.
- Boston Dynamics showcased the retiring hydraulic Atlas robot at CES, not the new electric version, which was anticipated but not demonstrated in a live, functional form.
- The new Atlas robot's design changes, such as a four-fingered hand and individual knee and ankle actuators, seem driven by cost reduction, simplicity, and manufacturability, rather than purely performance enhancements.
-
17:18 🤖 Humanoid robots are improving with design changes and AI integration, but still face challenges with reliability, mechanical stress, and instability.
- A discussion ensues about a humanoid robot's design, with Mario explaining its intended use on factory floors and reliability features, and Scott joking that it also succeeds as a comedy piece.
- As humanoid robots scale up production, reducing the number of SKUs and motor types may not be ideal, potentially leading to overactuation and mechanical stress issues, particularly in joints.
- Boston Dynamics' humanoid robot design changes, such as direct drive in the ankle and modified waist and knee joints, aim to address previous weaknesses and improve performance, particularly in movements like flipping.
- Google DeepMind partners with Boston Dynamics to integrate its AI capabilities with the Atlas humanoid robot, leveraging Hyundai's manufacturing scale and expertise.
- The discussion analyzes Boston Dynamics' humanoid robot, particularly a video of its impressive flip and landing, but notes that a sudden knee joint failure occurred, which was surprisingly recovered from by the robot's control system.
- The humanoid robot's design has limitations, such as decreased torque and mechanical advantage at certain points, which can cause instability and erratic movements, like suddenly shifting to one side.
-
28:08 🤖 Humanoid robots are becoming flashier with impressive abilities, but concerns around practicality, safety, and reliability remain, highlighting a gap between innovation and real-world application.
- Humanoid robots' abilities, such as performing backflips, are impressive but raise questions about their practicality, reliability, and potential for repeated failures on a factory floor.
- Humanoid robots are evolving from imitating human movements to creating their own motion primitives, unconstrained by human anatomy, but limitations in robot learning and safety considerations still hinder truly innovative designs.
- Humanoid robots still pose safety risks, such as falling or colliding with people, and current solutions, including sensors and warning signals, have limitations that need to be addressed.
- Alibaba's Genie Sim 3.0, an LLM-driven open-source simulation platform for embodied AI, offers a large dataset and high-fidelity environments, but its reliance on LLMs may limit its ability to generate truly novel scenarios, particularly for edge cases in robotics.
- Humanoid robots' simulations are improving with better visualization, but their physics engines still need significant improvement to enable robots to understand and interact with the physical world effectively.
- Plan for your retirement.
-
50:14 🤖 Humanoid robots may bring productivity gains, but their adoption is limited by economic and societal factors, and may also raise questions about human purpose and the meaning of life.
- We're approaching a technological singularity, but it's uncertain whether we'll slow down to navigate it safely or accelerate through it, risking unpredictable outcomes.
- Human society, built on scarcity, would be drastically changed if AI and robotics enabled people to not work and receive resources without effort, which wealthy individuals and governments might not want.
- The widespread adoption of humanoid robots that can do all the work for everyone will not magically result in a good life for all humans due to scarcity and economic limitations.
- Advances in technology and robotics will likely bring significant productivity gains and improvements in various fields, but may also raise questions about human purpose and the meaning of life as traditional notions of success and wealth are redefined.
- When considering a surgeon's qualifications, patients often want to know if the surgeon has experience with a particular procedure, essentially asking how many times they've successfully performed it.
- Humanoid robots may surpass human surgeons in precision and capability, but mass production constraints, such as material and supply chain limitations, will likely limit their widespread adoption.
-
01:00:21 🤖 Humanoid robots, despite advancing, still lack dexterity and capabilities to replace humans in complex tasks like surgery and patient care, and may take decades to potentially surpass human capabilities.
- Advances in AI and technology may change the future of education and medicine, but a medical doctor believes it will take decades, likely 40 years, for AI to potentially surpass human capabilities in surgery and other complex fields.
- Humanoid robots may excel in certain tasks, such as surgery and psychology, where precision and lack of human bias are beneficial, but human interaction and touch will still be valued in many areas.
- Humanoid robots in surgery, though advancing, still require human control and have not yet replaced doctors due to liability, training, and accessibility issues.
- Humanoid robots, like Optimus, may one day be capable of training human surgeons, but current robotic hands are not yet dexterous enough to perform surgery reliably.
- Humanoid robots currently lack the dexterity and interactive capabilities to replace humans in tasks that require complex hand movements and emotional connections, such as surgery and patient care.
- A humanoid robot's hand, costing around $50,000, impressively uses primarily two fingers and basic motions, raising questions about the necessity of a fully human-like hand.
-
01:18:15 🤖 Humanoid robots show incremental improvements in dexterity, but still struggle with fine motor skills and may rely on AI and proprioception rather than advanced senses to perform tasks.
- Humanoid robots' dexterity is more dependent on AI training and practice than physical hand capabilities, but current demos still lack true dexterity, struggling with tasks requiring fine motor skills and sensory feedback.
- Humanoid robots may not need highly advanced sense of touch to perform tasks, as proprioception and machine learning can enable robots to accomplish complex actions, such as writing and manipulating objects, with minimal sensory input.
- Humanoid robots may benefit from selectively disabling certain sensors, such as vision, and relying on other senses like proprioception or touch, to improve task performance in certain situations.
- Humanoid robots like Wuji and Shankai show incremental improvements in dexterity and design, but their development approaches and technical limitations, such as back drivability, reveal challenges in creating truly advanced humanoid robots.
- The humanoid robot "Matrix" bears striking similarities to Figure 3's design, including fabric and on/off switch, but with notable differences in actuation, sparking claims of potential copying.
- The U-joint's position seems odd as it doesn't appear to be kicked back enough, potentially limiting its flexibility due to insufficient offset between the tie rod and joint.
-
01:35:10 🤖 Humanoid robots are being showcased with impressive designs, but their functionality and practicality are questionable, with some developments prioritizing hype over viable products.
- The humanoid robot's design reveals its inner mechanisms, including exposed actuators for movement and a complex system of 25 tendons in the forearm for 27 degrees of freedom, suggesting a focus on flexibility and possibly cooling efficiency.
- The humanoid robot's design features a unique serial link mechanism in its leg, with a dedicated abductor and flexure, and its actuators appear surprisingly small, sparking questions about their power and torque density.
- The humanoid robot's hand design, with 5 tendons per digit and a potentially double-jointed wrist, allows for complex motions, but may still lack the nuanced coordination of human wrist and finger movements, such as combined extension, flexion, and radial deviation.
- Humanoid robots' wrist movements and functionality, crucial for tasks like hammering and using a screwdriver, are often overlooked in favor of gripping and finger movement.
- The presenter is skeptical about the capabilities of a humanoid robot showcased in a video, believing the design is overly simplistic and unrealistic, and criticizes the industry for releasing multiple versions of robots without deploying them in the market.
- Humanoid robot developments may prioritize creating a buzz for funding over functional products, with some companies showcasing prototypes that may not lead to viable end products.
- 01:51:44 🤝 The hosts wrap up a lengthy discussion, thanking viewers for joining in and looking forward to future chats.
-------------------------------------
Duration: 1:52:30
Publication Date: 2026-01-13T09:52:59Z
-------------------------------------