SuperMimic trains a humanoid whole-body motion tracker with one differentiable world model and reaches higher normalized return than SONIC using a single NVIDIA H100 GPU in three hours. It reduces the required GPU-hours by more than a thousand-fold compared with the 128-GPU, seven-day SONIC baseline.
Key points
129,514 motions trained; SuperMimic surpasses SONIC’s return with 1 GPU in 3 h.
SONIC fails 179 motions; SuperMimic fails none after 9.8 h on 2 GPUs.
Sources
- Original post
- SuperMimic - Linked page
- supermimic.github.io - Linked in the post
High-Precision RCM Motion for Ophthalmic Robotics achieves 0.142 mm residual
High-Precision RCM Motion for Ophthalmic Robotics achieves software-RCM residual below 0.142 mm and median tip error 32 µm. The paper presents a hierarchical-QP controller with authority-guided recovery planning and reports benchmark and physical arm results.
Key points
Benchmark: reaches 128/128 near-singular targets and 71/72 near-joint-limit targets (100% and 99%).
Physical test: worst-case kinematically reconstructed RCM residual 178 µm across four entry points.
Sources
Unitree G1 single policy for walking, climbing, crawling
ETH Zurich team trained a single policy for the Unitree G1 to walk, climb and crawl using 13.2 minutes of human motion data and reinforcement learning in simulation. The policy handles transitions and a ladder recovery sequence without explicit demonstration.
Key points
Trained on 13.2 minutes of human motion data to learn contact patterns.
Reinforcement learning in simulation adapts the policy to new terrain and enables a ladder recovery not present in the human demos.
Sources
China Mobile embodied intelligence team demo brain-controlled humanoid hand
China Mobile’s embodied intelligence team demonstrated brain-controlled grasping with a non-invasive BCI on a humanoid robot with a dexterous hand. The demo used AR glasses instead of a separate external screen.
Key points
Brain-controlled grasping using a non-invasive BCI system.
AR glasses used instead of a separate external screen.
Sources
Screw Actuator for Robot patent shows knee actuation
Screw Actuator for Robot patent describes a knee actuation method using a linear screw actuator inside the upper leg. The abstract lists the leg members, screw shaft, nut, motor, carrier and linkage arrangement.
Key points
Patent US20180172121A1 details a leg with upper and lower members coupled at a knee joint.
The design includes a screw shaft rotatable in a coaxial nut, a motor driving the shaft, a carrier mounted to the nut, and a linkage connecting the carrier to the lower leg.
Sources
Robotics events November 2026 and US humanoid firms
Robotics events November 2026 include multiple conferences in USA, Europe and China, plus a list of US humanoid robotics companies. The post lists dates, locations and company names.
Key points
Humanoid Robots Summit, Chicago, 3–4 November 2026
US humanoid robotics scene includes Tesla, Figure AI, Boston Dynamics, Apptronik, Agility Robotics, 1X
Sources
asylonrobotics combines drones and robot dogs for 24/7
asylonrobotics combines autonomous security drones and robot dogs to deliver 24/7 perimeter protection for critical infrastructure facilities. The approach targets continuous security of essential sites.
Key points
Combining autonomous security drones
Robot dogs for 24/7 perimeter protection of critical infrastructure facilities
Sources
- Original post
- hubs.la - Linked in the post