Capabilities
Each capability records who claimed it, whether a third party has reproduced it, and which benchmark the number came from. A self-reported benchmark is labelled as self-reported.
Tuesday, 5 August 2026 · 0 of 5 slots filled
01 · Developing
1X released a 25-degree-of-freedom tendon-driven hand for NEO. Low gear ratios keep every joint force-transparent and backdrivable, with tactile skin reading pressure and shear to catch a slipping object mid-slip. Demonstrated tasks run from picking screws and coins to plugging in a USB-C charger.
What would prove it: Third-party footage of the same tasks at production speed, repeated, with failure rate reported.
02 · Developing
FLUX-mimic — an early version of the FLUX 3 multimodal backbone — is reported running robots on Audi production lines, handling soft seals and cables that conventional automation could not, with 101ms reaction times on a single RTX 5090.
What would prove it: Confirmation from Audi, a line name, and a cycle-time figure.
03 · Developing
Robostral Navigate follows natural-language route instructions from a single RGB camera — no depth sensor, no LiDAR. Mistral reports 76.6% success on the unseen split of the R2R-CE benchmark, trained on roughly 2.4 million simulated trajectories across 350,000 scenes.
What would prove it: An independent run of the same benchmark, and any result at all outside simulation.
04 · Developing
Xiaomi-Robotics-1 pairs embodiment-free pre-training with a smaller real-robot post-training stage, trained on over 100,000 hours of real-world manipulation trajectories, and is pitched as a base model for household tasks.
What would prove it: A published evaluation, and disclosure of how the 100,000 hours were collected.
05 · Medium
Nvidia's humanoid blueprint pairs a Unitree body, five-fingered hands, onboard compute and software tools, aimed at academic researchers and slated to ship late 2026.
What would prove it: A shipping date met, and a bill of materials.