Welcome back, Thinkers! 🤖✨
This newsletter is the grand finale of our mini-series on robotics—now spotlighting the Top Humanoid Robots of 2025.
Previously in Think Stack 101:
Part I: From ancient automation dreams to Unimate, Shakey, and WABOT-2 (1984).
Part II: Post-1984 evolution—robots entering healthcare, homes, warehouses, and the 2020s humanoid surge.
Today: The humanoids that defined 2025—who moved fastest, handled tasks better, and looked closest to real deployment.
Humanoid robots have officially moved from “cool lab videos” to factory pilots, home prototypes, and early mass-production plans—especially as China pours serious money into embodied AI and manufacturing scale.
Let’s Explore.
Tesla - Optimus Gen 3
Tesla’s Optimus Gen 3 is positioned as the next iteration of Tesla’s general‑purpose humanoid worker—built to move from impressive demos toward real factory utility, leveraging Tesla’s computer vision and end‑to‑end AI stack as a key differentiator.
Early spec-style reporting around Gen 3 commonly describes a roughly human-scale platform (about 173 cm tall) with a lightweight build and more dexterous hands (often cited as 22 degrees of freedom per hand), aimed at safer, more practical manipulation tasks like carrying, sorting, and basic assembly work.
Optimus Gen 3 matters less for “flashy stunts” and more for Tesla’s intent to industrialize humanoids—tight integration of perception, control, and manufacturing scale, with an explicit goal of deploying in real workplace environments first.
Boston Dynamics - Atlas
Boston Dynamics’ Atlas in 2025 is best understood as an electric humanoid R&D flagship: a platform built to push the limits of whole‑body mobility and manipulation rather than to be a near-term mass-market product.
During 2025, Boston Dynamics highlighted Atlas policies trained with reinforcement learning and human motion references (motion capture), showcasing highly dynamic behaviors like walking, running, and crawling as the company matures learning-based control for real-world movement.
Atlas also signaled a shift from “stunts” to longer-horizon work skills: Boston Dynamics and Toyota Research Institute demonstrated Atlas running continuous manipulation-plus-locomotion task sequences using a Large Behavior Model approach, where one model can coordinate the entire body and adapt to unexpected mid-task changes.
Figure AI - Figure 03
Figure AI’s Figure 03 (revealed in 2025) is pitched as a ground-up redesign aimed at scaling a truly general-purpose humanoid—built to work with Figure’s on-board “Helix” system and learn tasks directly from people for use in the home and broader real-world settings.
Figure says Figure 03 was redesigned for manufacturability and deployment at scale, signaling a move toward repeatable real-world operation rather than one-off demos.
Agility Robotics - Digit v4
Agility Robotics’ Digit v4 is one of the most commercially oriented humanoid robots in 2025, designed specifically for warehouse and logistics work where robots must fit into human-built facilities.
In 2025 updates, Agility emphasized practical deployment features—like improved fleet operations, autonomous docking/charging, and interfaces that make it easier for on-site teams to monitor robot status—aimed at scaling Digit beyond pilot demos.
Digit’s positioning is clear: it’s not trying to win an acrobatics contest; it’s trying to become a dependable “shift worker” for repetitive material-handling workflows such as moving totes and operating around conveyors, racks, and carts.
Apptronik - Apollo
Apptronik’s Apollo in 2025 is best framed as a “practical humanoid” built for real work—especially repetitive intralogistics and manufacturing tasks in human-oriented facilities, rather than flashy acrobatics.
A key 2025 milestone was Apptronik’s collaboration with Jabil to scale production and validate Apollo inside real factory operations, with the pilot covering tasks like inspection, sorting, kitting, lineside delivery, fixture placement, and sub-assembly before deployments to customer sites.
The broader takeaway: Apollo’s narrative in 2025 centered on design-for-manufacture and scaling economics—using simplified, more affordable actuator designs and factory-grounded testing as the pathway from “humanoid demo” to deployable workforce robot.
Walker S2 (UBTECH)
UBTECH’s Walker S2 is positioned in 2025 as a factory-oriented humanoid designed for real shift work, with a headline feature: autonomous battery swapping aimed at near-continuous operation in industrial environments.
UBTECH highlights high-torque waist design and whole-body dynamic balance so Walker S2 can handle demanding motions like deep squats/stoop lifting and operate across a 0–1.8 m workspace while carrying up to 15 kg.
Walker S2 is less about viral agility and more about uptime + manufacturability—bringing humanoid form factors into production-line logistics, inspection, and material-handling scenarios where minimizing downtime is the differentiator.
Unitree G1 (Unitree)
Unitree’s G1 is a compact, developer-friendly humanoid that gained attention in 2025 for pushing a “humanoids for builders” angle—lighter, smaller, and more accessible than full-size industrial platforms, while still targeting real manipulation and mobility work.
In official materials, Unitree frames G1 around imitation + reinforcement learning and force/position hybrid control to enable more precise, compliant interaction with objects—important for research and early-stage deployment tasks.
G1 is also marketed as a flexible platform with a configurable joint count (commonly described in the 23–43 DOF range depending on configuration), which supports a wide spectrum of experiments from locomotion to dexterous hand work.
1X - NEO Gamma
1X’s NEO Gamma is a home-first humanoid designed around safe, compliant interaction—using tendon-driven actuation, a soft exterior, and whole-body control intended for everyday indoor spaces rather than factories.
In 2025 messaging, 1X emphasizes natural human-like mobility (walking with arm swing, squatting to pick objects, sitting on chairs) trained via reinforcement learning from human motion-capture data, plus a built-in language model for conversational control.
NEO Gamma’s published specs highlight practical home readiness signals—about 4 hours runtime (842 Wh), 1.4 m/s walking speed, and safety design choices like low-inertia tendon drives and reduced pinch points.
Sunday Robotics - Memo
Sunday Robotics’ Memo (revealed in 2025) is a home-first robot designed for everyday chores like clearing tables and loading dishwashers, intentionally using a wheeled base for stability instead of a biped humanoid form.
The company’s key differentiator is its data pipeline: Sunday says it trained Memo using large-scale, real household demonstrations captured with its “Skill Capture Glove,” collecting routines from 500+ homes (described as ~10 million episodes) to improve generalization in messy domestic environments.
Memo represents a clear bet that the fastest path to useful home robots is not acrobatics—it’s scalable real-world data + long-horizon task learning, with a “Founding Family” beta program announced for late 2026.
Phoenix Gen 8 (Sanctuary AI)
Phoenix Gen 8 is Sanctuary AI’s 2025 iteration of its general-purpose humanoid program, optimized less for “flashy biped demos” and more for capturing high-quality training data that improves real-world task performance.
A defining Gen 8 change is the move to a wheeled base, which Sanctuary says was driven by customer feedback that biped legs can be too fragile for the strong, stable torso needed for precise, safe, useful work.
Sanctuary also highlights upgrades to telemetry plus depth/vision cameras (wider field of view and higher resolution) and a more manufacturing-friendly design aimed at simpler builds and lower bill-of-materials cost.
Engineered Arts - Ameca (Generation 3)
Engineered Arts’ Ameca (Generation 3) in 2025 stands out as a flagship “social humanoid” focused on high-fidelity human–robot interaction—natural conversation, eye contact, and expressive facial motion—more than heavy-duty industrial work.
Engineered Arts positions Ameca as a programmable engagement platform (events, museums, research, customer experiences), with detailed published specs and a modular system approach that supports ongoing upgrades.
Ameca’s core differentiator remains its realism in communication: cameras, microphones, and a highly articulated head/face system designed to deliver lifelike expressions and responsive interaction in real time.
Sophia (Hanson Robotics)
Sophia (Hanson Robotics) remains one of the most recognizable social humanoid robots in 2025, designed primarily for human–robot interaction rather than physical labor.
Hanson Robotics positions Sophia as a platform that blends robotics engineering and expressive design, using lifelike facial expressions and conversation to support public engagement, education, and AI/HRI research.
Sophia’s value is less about warehouse productivity and more about being a long-running, high-visibility benchmark for how believable face-to-face interaction with robots can feel in real public settings.
Pepper (SoftBank)
Pepper is SoftBank Robotics’ iconic social service robot, still widely referenced in 2025 as an early benchmark for customer-facing humanoids in retail, hospitality, healthcare, and education.
Rather than doing heavy physical work, Pepper’s strength is interaction: it combines a 10.1-inch chest tablet with multi-modal sensing (cameras, microphones, depth sensing, touch sensors) to greet people, guide visitors, and run scripted or semi-scripted engagement flows.
Pepper represents the “service UX” branch of humanoids—proving that deployment value can come from communication, wayfinding, and workflow integration even without advanced manipulation or factory-grade strength.
NEURA Robotics – 4NE-1
NEURA Robotics’ 4NE‑1 emerged in 2025 as one of Europe’s most ambitious “cognitive humanoid” pushes—designed to work safely alongside people in both industrial and service environments, not just in lab demos.
In mid‑2025 coverage around its third‑generation unveiling, NEURA emphasized an “artificial skin” concept for safer human‑robot collaboration and a dual‑battery design aimed at 24/7 operation, positioning 4NE‑1 around uptime and close-proximity work.
NEURA also tied 4NE‑1 to scalable manufacturing via its “NEURA Hive” production method—framing the robot as a platform meant for volume production, not a one-off prototype.
Engine AI - T800
Engine AI’s T800 made waves in late 2025 as a full-size, high-power humanoid designed for industrial-grade physical performance, with marketing that emphasizes impact tolerance and “combat-style” stability testing.
According to EngineAI’s published product details, T800 is 173 cm tall, uses a 29‑DOF body design (excluding dexterous hands), targets up to 450 N·m maximum joint torque, and lists hardware-supported movement speed of at least 3 m/s.
EngineAI also highlights full leg-joint active cooling for sustained high-intensity operation (up to 4 hours) plus a modular battery architecture and stereo vision + LiDAR perception stack—positioning T800 for logistics, factory work, and other demanding environments.
Unitree H2 (Destiny)
Unitree’s H2 (“Destiny Awakening”) is the company’s full-size flagship humanoid in 2025, positioned as a step up from earlier models with more human-like proportions and smoother, more expressive whole‑body motion.
Official specs highlight 31 degrees of freedom, up to 360 N·m maximum leg-joint torque and 120 N·m maximum arm-joint torque, plus wide‑FoV binocular vision and a bionic head concept aimed at more natural interaction.
H2 signals Unitree’s shift from “affordable humanoid access” to a more capable, premium platform meant for both enterprise pilots and research—where agility, perception, and upper-body control matter as much as raw walking speed.
Robot Era – STAR 1
Robot Era’s STAR 1 drew attention in 2025 for prioritizing raw biped speed and outdoor robustness, positioning itself as a “performance humanoid” rather than a slow, lab-only prototype.
Robot Era and third-party coverage cite STAR 1 at about 171 cm tall with 55 DOF, capable of reaching roughly 3.6 m/s running speed, backed by high-torque actuation (reported up to 400 N·m) and an AI compute figure often cited around 275 TOPS.
STAR 1 is a strong signal that the humanoid race is now being pushed by measurable athletic metrics—speed, stability across mixed terrain, and repeatable dynamic locomotion—alongside dexterity.
DEEP Robotics – Dr01
DEEP Robotics’ DR01 (Dr.01) is positioned in 2025 as an “embodied intelligence explorer” humanoid built for messy, real-world mobility—prioritizing stability, recovery, and outdoor/industrial terrain traversal over polished indoor-only demos.
The company highlights locomotion across uneven ground (including ~25° slopes and ~18 cm steps), balance recovery under disturbances (like pushes or slippery surfaces), and a “fusion perception” approach that combines robot-state sensing with environmental perception for adaptive control.
DR01 is a strong signal that the humanoid category is expanding beyond factories into inspection, response, and other environments where terrain robustness and fast self-recovery are the real differentiators.
Kepler Robotics – Forerunner K2
Kepler Robotics’ Forerunner K2 (often nicknamed “Bumblebee”) is a factory-leaning humanoid positioned around endurance and shift-ready economics in 2025, aiming to be a practical “blue-collar robot” for logistics and manufacturing environments.
Reports highlight a 52‑DOF full-body design, rope-driven tactile hands rated up to 15 kg per hand with dense fingertip contact sensing, and a 2.33 kWh battery targeting up to 8 hours of operation with flexible charging options (including automatic charging).
K2’s differentiator is less about viral motion and more about deployment fundamentals—maintainability (including a star-shaped wiring layout), long runtime, and manipulation designed for repetitive industrial handling.
Dobot Robotics – Dobot Atom
Dobot Robotics’ Dobot Atom is a precision-first humanoid that stood out in 2025 for targeting real industrial manipulation, not flashy athletic demos. It’s marketed around industrial-grade ±0.05 mm repeatability and a 41‑DOF full-size configuration (Atom Max), positioning it for delicate tasks like calibration and electronics-style handling.
A key differentiator is Atom’s efficiency-focused mobility (including a straight-knee walking concept highlighted across industry coverage), paired with a perception stack that includes Full HD vision and Intel RealSense depth sensing.
Astribot – Stardust Smart S1
Astribot’s Stardust Smart S1 stood out in 2025 as a “dexterity-first” humanoid platform, marketed around extremely fast arm motion and human-like upper-body operation aimed at research and real-world task learning.
Astribot lists 7 DoF per arm, 5 kg payload per arm (horizontal reach), and end-effector speeds of ≥10 m/s, positioning S1 for high-speed manipulation tasks where smoothness and precision matter more than heavy lifting.
Keenon Robotics – XMAN-R1
Keenon Robotics’ XMAN‑R1 is a service-focused humanoid introduced as part of Keenon’s push into “multi-form” commercial robotics in 2025, designed to interact with people and coordinate tasks alongside Keenon’s existing delivery and cleaning robot lineup.
Instead of targeting factory strength or acrobatics, XMAN‑R1 is framed around practical hospitality and retail workflows—think guiding, carrying, delivery/collection, and orchestrating tasks in shared spaces with other robots.
PND Robotics – Adam Lite
PND Robotics’ Adam gained visibility in 2025 as a developer-oriented humanoid platform that blends lifelike motion with performance-style demonstrations (notably public showcases where Adam appeared on stage and performed coordinated routines).
Coverage around Adam in 2025 emphasizes a modular architecture and learning-based control (imitation learning + reinforcement learning), positioning it as a flexible base for industrial pilots, research labs, and content-friendly demos that prove balance and whole-body coordination.
Engine AI – SE01
Engine AI’s SE01 is a full-size humanoid that drew attention in 2025 for its notably natural, human-like walking gait, with EngineAI positioning it as a versatile platform for research, education, and industrial scenarios rather than a single-purpose factory robot.
SE01 is described as using an end-to-end neural-network approach for locomotion (combining reinforcement and imitation learning) and pairing that software stack with dual NVIDIA + Intel processors plus high-precision stereo vision for real-world perception and navigation.
HMND 01 Alpha Bipedal
HMND 01 Alpha Bipedal is a UK-built, industrial-oriented humanoid revealed in 2025, designed to extend automation into “brownfield” spaces where fixed automation and wheeled robots struggle (stairs, tight aisles, irregular layouts).
In its bipedal configuration, HMND 01 Alpha is described as human-scale (179 cm) and ~90 kg, with 29 DOF (excluding end-effectors), ~3 hours runtime, and ~15 kg bimanual payload—framing it as a practical factory helper rather than a stunt robot.
A headline claim around the platform is rapid bring-up: the company says the robot was “walking within 48 hours of assembly,” backed by extensive simulation training and reinforcement-learning-based push recovery for stability.
Caltech X1 Robot
Caltech’s X1 in 2025 is best described as a multi-robot response system rather than a single humanoid: it pairs a modified Unitree G1 humanoid with Caltech’s “M4” transforming robot that can both fly and drive after launching from the humanoid’s back.
Caltech frames X1 as a mobility-and-autonomy concept for emergency response, demonstrating coordinated walking + drone deployment + ground driving to traverse complex campus terrain more efficiently than any single robot modality alone.
LimX Dynamics - Oli
LimX Dynamics’ Oli emerged in 2025 as a full-size, developer-focused humanoid aimed at accelerating embodied-AI research into real deployments, with a modular hardware/software architecture and an open SDK for perception, motion control, and task orchestration.
LimX states Oli stands 165 cm tall and has 31 active degrees of freedom (excluding end effectors), and is offered in multiple editions (Lite/EDU/Super) targeted at researchers, developers, and integrators.
Oli is positioned less as a polished “consumer humanoid” and more as a configurable platform—supporting different end-effectors (e.g., grippers or dexterous hands), third-party sensors, and over-the-air motion/controller updates to speed iteration.
XPENG IRON
XPENG’s IRON is an automaker-born humanoid that, in 2025, is framed as a bridge between EV-grade autonomy tech and embodied robotics—built for real spaces like factories and retail before attempting the home.
Publicly listed details describe IRON as roughly 1.73 m tall and ~70 kg, with ~60 joints and ~200 DOF, plus 22‑DOF hands and “720°” vision positioning for high awareness in cluttered human environments.
IRON is a strong signal that OEMs are entering humanoids with a serious advantage—manufacturing discipline and mature perception stacks—aiming to iterate quickly by deploying the robot inside their own operations.
TARS Robotics
TARS Robotics emerged in 2025 as a fast-moving Beijing-based embodied-AI startup focused on building a tight data–model–robot loop for real-world manipulation.
In late 2025, TARS showcased humanoid robots performing delicate “hard to automate” tasks like hand embroidery, and described its stack as SenseHub (human-centric data capture) feeding an embodied model it calls AWE 2.0, which then drives execution on its T- and A-Series robots.
The company also drew attention for unusually large early funding and an aggressive roadmap, positioning itself as a serious contender in China’s next wave of humanoid/embodied intelligence teams.
Agile Robotics - Agile One
Agile Robots’ Agile ONE is a factory-first humanoid unveiled in 2025, positioned around “Physical AI” for manufacturing and logistics—moving between stations, handling parts, and using tools in existing industrial workflows.
Its standout feature is dexterity: Agile Robots says Agile ONE’s hands use modular fingers with 21 joints and integrated fingertip plus force/torque sensing, aiming to handle both delicate manipulation and firm grasping.
Agile Robots also states Agile ONE can walk at up to 2 m/s, reinforcing the pitch that it’s designed to keep pace with human operations on the shop floor.
China’s “robot school”
China’s “robot school” story in 2025 centers on Beijing’s Phase II Humanoid Robot Data Training Center, a facility that simulates full-scale factory lines and home environments so humanoid robots can practice tasks like sorting, packing, cooking, and room organization before deployment.
The key idea is data: the center is designed to generate millions of high-quality, standardized training records annually, addressing a major bottleneck in embodied AI—getting enough consistent real-world task data for robots to generalize beyond demos.
It’s also structured like education: robots choose “majors” (manufacturing, smart home, elderly care, 5G-integrated scenarios) across 16 disciplines, train with human instructors in repeated trials, and some “graduates” are already reported working in roles like factory material handling and power inspection.
Why 2026 can be even bigger 🚀
2025 felt like the year humanoids graduated from “stage demos” to “real environments.” We saw full-time commercial work (like Digit’s ongoing warehouse deployment), dedicated manufacturing scale-ups (RoboFab), and a clear push toward home + general-purpose designs (Figure 03). Even the legacy leaders doubled down on “real-world applications,” with the fully electric Atlas positioned for practical industry work.
2026 could be the year pilots start turning into repeatable playbooks: more robots in more sites, better autonomy driven by real-world data (even if many systems still need supervision), and louder “commercialization” announcements on big stages—like Boston Dynamics’ next-gen Atlas debut at CES. The opportunity is huge, but so is the responsibility: safety, privacy, and security will decide how fast this wave becomes everyday reality.
Drop a comment or reply to this newsletter – we’d love to hear your perspectives!
Think Stack 101 — Your Daily Stack of Knowledge, Wisdom, Ideas 🧠
Cut through the noise!
Follow Think Stack 101 on WhatsApp, LinkedIn, Telegram, and Substack for your daily stack of knowledge, wisdom, ideas, and innovation to power up your thinking. 🚀🧠✨
✨ One smart thought a day. Just the boost your mind needs.
So here we are! Rise of the Machines. They seem relatively deliberate (slow?) but I suspect they will gradually become quicker (with superior dexterity) than any human. I can see cause for both celebration and concern…