ACQUIRING TARGETS...
· DroidHunt news
MIT professor Pulkit Agrawal and ex-Google DeepMind scientist Tuomas Haarnoja's startup Eka Robotics has raised USD 150M at a USD 1.45B post-money valuation, led by Andreessen Horowitz and F-Prime. Eka's edge is a vision-force-action (VFA) architecture — a force-sensitive gripper feeding a deep-learning model, an alternative to the dominant vision-language-action (VLA) paradigm. The company claims ~1,500 object picks per hour (up to 1 kg each) vs ~600 for a human, at >95% success. Gripper hardware costs under USD 1,000 and is rated for 5M duty cycles (goal: 20M). RL training runs at scale in physics simulation — "thousands of hours of robot experience in minutes." Secretive industrial pilots are under way in pick-and-place and wire manipulation. Single-outlet exclusive; company figures are self-reported.
Why it matters: A USD 1.45B mark on a Series A with no public product shows capital is pricing embodied-AI teams on model architecture, not deployments. Force feedback as the control signal is a genuine bet against the VLA consensus, with clear benchmark stakes if the pilot numbers hold.
Sources: techcrunch.com
Tags: eka-robotics funding us reported
A preprint (arXiv 2610.28480, Qiayuan Liao et al.) introduces SuperMimic, a motion-tracking method that amortizes learning across all motions via a single differentiable world model of shared dynamics, fitted from a rigid-body simulator, and trains by backpropagating through imagined rollouts. On a benchmark of 129,514 motion clips, the authors report 25 failures vs 179 for a strong baseline trained for 7 days on 128 GPUs — while SuperMimic reaches better performance in 3 hours on a single GPU, a >1,000x reduction in GPU-hours — and zero failures after 9.8 hours on two GPUs. The policy transfers to a physical Unitree G1, including acrobatic motions. Results are self-reported and not yet peer-reviewed.
Why it matters: Whole-body humanoid motion tracking has been compute-gated, which made large-scale scaling-law studies impractical. Cutting training to one GPU-hour territory democratizes the capability and makes motion data collection — the bottleneck for real-world humanoid skill — dramatically cheaper.
Sources: arxiv.org
Tags: supermimic research us reported
← Newer: Oct 22, 2026 Older: Oct 20, 2026 →
Home Every digest Robot catalog Drones on DronePriceHunt Privacy & affiliate disclosure RSS