Humanoids Sprint Toward Humans, AI Eyes Post-Transformer Era
Robotics hits athletic peaks with 12km/h sprints and 96.5% tennis rallies; Altman predicts transformers' replacement by AI-designed architectures, enabling AGI in 2 years.
Humanoids Achieve Near-Human Athleticism and Dexterity
China and South Korea lead humanoid breakthroughs, pushing speed, sports skills, and manipulation toward human levels. KIST's V0.7 humanoid (75kg, 5'5") runs 12km/h on flat ground, jumps 30cm steps, and performs soccer drills plus moonwalks. Built in-house with quasi-direct drive motors (knee: 320Nm torque), high-torque low-ratio gearboxes, and deep RL trained on human motion data, it uses proprioception for uneven terrain without cameras. Future targets: 14km/h, 40cm steps, ladder climbing. Unitree G1, trained via Leighton's latent action space on 5 hours of amateur tennis data, hits 96.5% rally success over 10,000 trials from fore/backcourt, blending RL and simulation for dynamic sports like soccer or parkour.
Speed claims escalate: Unitree's Bolt reaches 10m/s (near Usain Bolt's 10.44m/s average), with founder Wang Xingxing predicting sub-10s 100m sprints by mid-year. Challenge remains generalization—controlled demos falter in unpredictable environments. Hands advance too: Tasbot's DG5FS (20 DoF, 880g, back-drivable joints for safe impacts) and Samsung's tendon-driven tactile hands target dexterous manipulation. Market for five-finger hands projected at $876M by 2030.
"Humanoid robots may soon rival or even beat the fastest human ever in sprinting." — Wang Xingxing, Unitree founder
Exotic Robotics Tackle Endurance, Sustainability, and Safety
Non-humanoid innovations address deployment hurdles. Cranfield's Wanderbot uses wind-powered Savonius turbine and Jansen linkage for battery-free movement (20% typical energy drain), ideal for deserts/planets; 3D-printed for on-site repairs, low TRL but eyed for space. NUS's Ostrobot, fish-inspired with lab-grown antagonistic muscles, self-trains to 467mm/min swim speed (3x standard), 7.05mN force—controlled via electricity/sound.
Safety failures highlight real-world gaps: Agibot X2 at hot pot restaurant swung erratically, smashing dishes near boiling soup—blamed on guest proximity, underscoring demo-to-deployment risks. Counter: Oklahoma State's neuradaptive system reads EEG error-related potentials (ERPs) via cap, adapting in ms for nuclear/deep-sea tasks; uses NVIDIA Isaac Lab/Sim, signal temporal logic for rules, personalizing to user brains—extends to prosthetics.
Sustainability: Seoul Nat'l U.'s compostable soft robot (PGS elastomer) endures 1M cycles, biodegradable electronics/sensors (curvature, strain, pH); decomposes tox-free in months. Production scales: UBTech-Seamens deal targets 10k units/year by 2026, leveraging digital sim/manufacturing amid 1.4B yuan orders.
"Robots that look great in controlled demos can become a problem fast in crowded, unpredictable, real-world spaces."
AI Architectures and Capabilities Signal Paradigm Shifts
Sam Altman declares transformers (ChatGPT's backbone) inefficient for long contexts—10x length demands 100x compute—and ripe for replacement, akin to transformers over LSTMs. AI aids discovery, accelerating loops toward AGI in 2 years, programming agents as next boom (one-person companies, AI CEOs). Mamba exemplifies efficient alternatives. Early OpenAI: apartment origins, rapid ideation.
Apple's Leto reconstructs 3D objects from one image with consistent lighting/reflections; trained on 150-view/3-light objects, compresses to latent rep then reconstructs. Inspio World FM builds real-time 3D spatial understanding (RTX 4090) via multi-view consistency, anchors/implicit memory—key for robotics stability.
Agents act: Manus' My Computer controls local PCs (files, CLI, GPU) with permissions. Others: Mistral's Leanscroll self-fixes code; Zhipu GLM5 Turbo executes workflows.
"The transformer architecture, the thing that powers ChatGPT and most modern AI, is not the final step." — Sam Altman
"Current AI models are already smart enough to help discover that next architecture." — Sam Altman
Key Takeaways
- Train humanoids with RL + imperfect human data (e.g., 5h tennis) via latent spaces for 96.5% dynamic task success; simulate hardware mismatches precisely.
- Prioritize generalization over demo speed—test in unpredictable settings early.
- For endurance, explore wind/Jansen linkages or self-training bio-muscles to cut battery reliance.
- Integrate EEG/ERPs for human-robot safety loops in high-risk ops; personalize decoding models.
- Scale production with digital twins (UBTech model) before humanoid hype turns industrial.
- Bet on post-transformer efficiency (Mamba-like); use AI to co-design architectures.
- Build 3D-consistent models (Leto/World FM) for robotics perception; run real-time on consumer GPUs.
- Deploy local agents (My Computer) for action over chat; gate with permissions.
- Prototype compostable materials (PGS) now to preempt robotics e-waste at scale.