Robot hands are getting much more serious. Unitree’s Dex5-1 has 20 DoF, with 16 active joints and 4 coupled joints. It weighs 1.1 kg, reaches ±1 mm fingertip repeatability, 10 N fingertip force and runs control at 1000 Hz. The Dex5-1P version adds 94 pressure sensors across the hand. Linkerbot’s L30 uses 18 active + 4 passive DoF with tendon drive. It weighs 1.192 kg, opens and closes in 0.2 s, reaches 39 N grip force, ±0.2 mm repeatability and a listed 15 kg max load. Communication runs at 500 Hz over CAN FD. SharpaWave has 22 active DoF, 150 N grip force and up to 20 N at the fingertip. Its tactile system reaches 240×240 resolution at up to 180 fps, with ±1 mm fingertip repeatability. I keep coming back to the sensing numbers. Unitree is packing 94 pressure sensors into one hand, while Sharpa is pushing dense tactile data at 180 fps and Linkerbot is targeting a 0.2 s open-close cycle. Three very different approaches to giving humanoids better hands.

LinkerBot is building robot hands that can move much closer to a human hand than a standard gripper. These clips show its dexterous-hand work across small-part handling, food manipulation finger coordination and grasping objects with very different shapes. One of LinkerBot’s current research hands the Linker Hand O30, has 20 fully active degrees of freedom, with all 20 joints controlled independently. It weighs 730 g, supports a rated payload of 30 kg, and produces up to 70 N of five-finger grip force. The O30 uses direct drive, communicates at 500 Hz through CAN or CAN FD, has ±0.20 mm repeat positioning accuracy, and opens or closes in 0.8 seconds. LinkerBot also specifies 24 N at the thumb tip and up to 30 N at each of the other fingertips. The food handling caught my attention most. An avocado, dough or other irregular object forces the hand to coordinate several contact points without crushing or dropping it. LinkerBot is also developing linkage driven and tendon-driven hands alongside the direct-drive O30, giving researchers several mechanical approaches for dexterous manipulation.

Speech recognition gets much harder the moment people start talking like they actually do. People switch languages mid-sentence, restart thoughts, use weak microphones and speak over traffic or room noise. I’d test Hojo-ASR-Multi-V1 first on English and French code-switching, then add background noise and phone-recorded audio. That should expose where the transcription starts to break.

Most people have already heard about Tiangong Omni but many still don’t know what this small humanoid can actually do. Here’s what you should know. Tiangong Omni was built by the Beijing Humanoid Robot Innovation Center, also known as X-Humanoid. The flagship version stands 1.35 m tall and weighs 39.5 kg. It has 31 active DoF, with 6 joints per leg, 7 per arm, 3 in the waist and 2 in the neck. Inside is an AI computer rated at 2,070 TFLOPS. Its sensor stack includes a depth camera LiDAR and a high-precision IMU. It can walk run climb stairs, cross sparse footholds handle uneven terrain recover after falls and perform mobile manipulation. The footage here shows some of those abilities under much harder conditions. At the 2026 World Humanoid Robot Games Tiangong Omni completed a 400 m obstacle course containing 16 different obstacles and won gold in 4:11.44. It also won the small-size 400 m race in 45.66 seconds. The locomotion system used in the obstacle competition came from the SOLO research project. The policy runs at 50 Hz with one chest-mounted depth camera and proprioception, controlling 25 joints without external localization or motion capture. SOLO also completed a continuous 1.5 km outdoor route across stairs, slopes, grass and uneven ground What caught my attention in this clip is the recovery. Omni can hit the ground, reorganize its body, stand back up and continue toward the next obstacle. That ability becomes very useful once humanoids leave flat lab floors

SkipVLA takes a pretty practical approach to VLA robot control. They tested it on a 6-DoF YAM arm. During free-space motion, the arm follows a classical planner. The VLA is called around grasping and placing. A small scorer looks at frozen vision-language features and picks the next 3D target pose. They did this without collecting extra demonstrations for that module. The real robot results are solid. With π0.5, three-block stacking went from 14/20 to 17/20 successful runs, while average completion time dropped from 47.62 s to 36.13 s. With MolmoAct2, placing two blocks in a box went from 18/20 to 20/20. Time dropped from 76.04 s to 41.66 s, and Jetson Thor compute energy fell from 3,237 J to 1,562 J. The largest measured time reduction was 59.6%. The stacking result caught my attention most. The robot finished more trials while calling the VLA less during the easy free-space parts of the motion. There is still a limitation. The switching rule currently depends on gripper open and close events, so the paper only tests pick-and-place style tasks.

This might completely change the game for humanoid robots. AGIBOT A3 joined celebrities on stage at the Greater Bay Area Film Concert in Macao. Powered by WITA-Omni. A3 followed the conversation in real time while coordinating speech with body movements and expressions. WITA-Omni uses a Thinker-Talker-Actor architecture. The Thinker processes text images audio and audiovisual input inside a shared representation for reasoning and interaction decisions. The Talker generates speech from that shared state. The Actor generates physical movements and facial expressions alongside the speech. The system can keep observing its surroundings and receiving audio while preparing a response. That shared timeline helps connect what the robot sees and hears with what it says and physically does next. You can see it gesture with both arms. Change its stance. React to the people around it. Then respond at the right moment during the conversation. I kept watching the timing between its voice and body motion. Live human interaction gives a robot much less room for scripted timing. @AGIBOTofficial #AGIBOT# #AGIBOTA3# #WITAOmni# #HumanoidRobotics# #EmbodiedAI#

PrimeBOT just held its dedicated Q1 and T1 robot launch in Shanghai. The Q1 is an 88 cm humanoid with 22 DoF and full-body force control according to the company. The T1 takes a different route. It can switch between a wheeled humanoid body for indoor movement and a quadruped body for stairs, slopes and outdoor terrain. PrimeBOT also gives it person following, autonomous filming and 15+ camera movement presets. I keep coming back to that body design. A wheeled base saves energy on flat floors, while four legs give the same robot another option when the ground stops being flat. These robots were shown before today. The fresh update is the September 20 consumer launch event in Shanghai. The performance figures still come from PrimeBOT.

I gave GPT-6 Astra this image and asked it to recreate the robotic arm. This is what it came up with. You get a good look at the full mechanical chain from the rotating base to the shoulder, elbow, wrist and gripper. The circular joint modules sit between the slotted links, and the end effector uses a compact linkage system to drive both fingers around the object. What I kept looking at was the gripper. You can almost trace the motion from the actuator all the way to the fingertips. Nice reference for anyone studying robot arm design, gripper linkages and basic kinematics. How close do you think it got?

A Unitree G1 just showed how fast humanoid policies can fail when the environment moves outside their training range. New work from the University of Florida trained a G1 to perform roofing-style motions on pitched surfaces using human VR demonstrations. The human wore a PICO headset, two controllers and two ankle trackers. Those motions were retargeted to the 29-DoF G1. The team also measured the roof geometry and corrected foot support, hand clearance and body penetration before PPO training. In simulation, the robot performed nailgun positioning, hammering and lateral pushing. Work-clearance error reached 0.531 cm for nailgun positioning, 0.256 cm for hammering and 0.424 cm for pushing. All three tasks finished 3/3 successful runs. The 25° roof test is the result I keep coming back to. A policy trained on 9° and 17° slopes failed all 10 trials on an unseen 25° roof. Once 25° examples were added to training, the same setup completed 10/10 trials at 9°, 17° and 25°. The policies were also transferred to a physical G1. Base-frame motion error was 27.6 mm during uphill walking, 60.2 mm for nailgun positioning, 79.9 mm for hammering and 61.4 mm for bending. The hardware tests used a safety hoist. No nails were fired and hammer impact was not measured. Roof edges, shingles and weather were outside the test setup. That 10/10 to 0/10 drop at an unseen slope is a very concrete look at how narrow locomotion generalization can still be for humanoids.

50 websites for anyone who spends too much time following humanoid robots. I put this list together across the full stack Robot companies and platforms Research papers Datasets and open source Simulation and robot learning Industry news and funding You’ll find Figure, Unitree, AGIBOT, Boston Dynamics, arXiv, LeRobot, MuJoCo, Isaac Lab, ROS, The Robot Report and many more. I like seeing all five categories on one page because following humanoid robotics usually means jumping between company releases, papers, datasets and simulation tools every day. Save it. You’ll probably come back to a few of these.

Sharpa Robotics just dropped a new hand video and the level keeps going up. This is the Sharpa Wave running WM Craftnet on a human scale fivefinger hand with 22 active DoF. The policy combines wrist depth, tactile sensing, proprioception and previous actions. The hand can rotate different objects in-hand, recover after external pushes and continue manipulating objects it was never trained on. The numbers are strong. 175/200 successful real-world rotation trials across 20 objects. A world-model prior trained on 9 objects was transferred to 49 new objects. Fall rate went from 6% to 0.3%. The Wave hardware itself has 22 actuators, up to 20 N fingertip force, 240×240 tactile sensing at up to 180 fps and 0.02 N pressure sensitivity. What caught my attention is the recovery behavior. The fingers keep changing contact points after the object slips or gets pushed instead of replaying the same finger motion. That is the kind of dexterity I want to see more of in robotic hands. According to Sharpa’s current specifications: • DTA tactile sensors on the fingers with a resolution of up to 240 × 240 • Pressure detection • Slip detection • Force change detection • Contact point localization • 6-axis force and torque measurement: Fx, Fy, Fz, Mx, My, Mz • Tactile sensing at up to 180 fps • 20 ms reported latency • Force detection range from 0 to 30 N • Maximum sensor load of 50 N • Sharpa also describes a miniature camera integrated into each fingertip for visuo-tactile sensing.

Sharpa Robotics just dropped a new hand video and the level keeps going up. This is the Sharpa Wave running WM Craftnet on a human scale fivefinger hand with 22 active DoF. The policy combines wrist depth, tactile sensing, proprioception and previous actions. The hand can rotate different objects in-hand, recover after external pushes and continue manipulating objects it was never trained on. The numbers are strong. 175/200 successful real-world rotation trials across 20 objects. A world-model prior trained on 9 objects was transferred to 49 new objects. Fall rate went from 6% to 0.3%. The Wave hardware itself has 22 actuators, up to 20 N fingertip force, 240×240 tactile sensing at up to 180 fps and 0.02 N pressure sensitivity. What caught my attention is the recovery behavior. The fingers keep changing contact points after the object slips or gets pushed instead of replaying the same finger motion. That is the kind of dexterity I want to see more of in robotic hands. According to Sharpa’s current specifications: • DTA tactile sensors on the fingers with a resolution of up to 240 × 240 • Pressure detection • Slip detection • Force change detection • Contact point localization • 6-axis force and torque measurement: Fx, Fy, Fz, Mx, My, Mz • Tactile sensing at up to 180 fps • 20 ms reported latency • Force detection range from 0 to 30 N • Maximum sensor load of 50 N • Sharpa also describes a miniature camera integrated into each fingertip for visuo-tactile sensing.

50 websites for anyone who spends too much time following humanoid robots. I put this list together across the full stack Robot companies and platforms Research papers Datasets and open source Simulation and robot learning Industry news and funding You’ll find Figure, Unitree, AGIBOT, Boston Dynamics, arXiv, LeRobot, MuJoCo, Isaac Lab, ROS, The Robot Report and many more. I like seeing all five categories on one page because following humanoid robotics usually means jumping between company releases, papers, datasets and simulation tools every day. Save it. You’ll probably come back to a few of these.

Two very different projects. Same kind of engineering instinct. The upper clip uses a cockroach with a very minimal mechanical setup. The lower one takes a quadcopter and surrounds it with tracked frames so the same machine can move on land, travel through water and take off into the air. What caught my eye is how both designs start from a physical behavior instead of adding layers of software. For the air land water drone the engineering tradeoff is especially interesting. Flying helps with obstacles and speed. Driving can save battery when the terrain allows it. Water mobility gives the same platform another route when roads disappear. Of course, every extra mobility mode adds weight, moving parts and harder control transitions. Still I like this direction. Build the machine around the environment it has to cross. I’m still wondering what the cockroach project in the top clip is actually meant to do though. Does anyone know the intended use case?