§1 Cover · The Resistance · Field Report #0451

Is a Cat Smarter Than AI?

Finding: In the physical world, yes. Subject Felis catus outperforms every known machine on a 35-gram brain[1] drawing roughly a third of a watt.[2]

AI won at chess. It still can't land like a cat. Each ability below falls on its own Judgment Day.

Judgment Day status: 0 / 12 abilities matched

§2 Spec sheet · Subject vs. machine

Hardware comparison

SpecF. catusBest available machine
Brain 34.9 g, 1.2 billion neurons[1] NVIDIA Jetson Thor module, current robot “brain” of choice[5]
Brain power ≈0.3 W[2] (est.) up to 130 W[5]
Whole-body power ≈13 W, all-inclusive[3] ≈380 W average (Spot: 564 Wh over ~90 min)[4]
Runtime 14 years (median)[6] ~90 minutes per charge[4]
Fuel Chemical (food), ≈274 kcal/day Li-ion; grid recharge
Training data Single agent, own experience, weeks ~130,000 demos, 13 robots, 17 months (RT-1)[7]
Self-repair Native tissue repair Human intervention still essential[8]

Brain power ratio, cat vs. robot computer: about 1 : 400. Cat figure includes locomotion control.

§3 The list · Abilities machines still cannot match

Outstanding capabilities

Grouped by sector. Expand an entry for the evidence and the exact condition under which it gets struck off.

Actuation & Manipulation

  • ☐Muscle-grade actuation[9]

    Evidence: Survey of 13 artificial-muscle classes (SMA, dielectric elastomer, IPMC, twisted yarns, pneumatic, electrostatic…) and their limitations against biological muscle. Production robots still use electric motors with gearboxes.

    Struck when: A single actuator technology matches mammalian skeletal muscle on power density, efficiency, strain and cycle life simultaneously, and ships in a commercial robot.

  • ☐Dexterous in-hand manipulation[10]

    Evidence: In-hand manipulation, deformable objects and contact-rich tasks remain open problems despite decades of hardware and software work.

    Struck when: Commercial robot reliably reorients and uses arbitrary unseen objects, rigid and deformable, at human-comparable speed.

  • ☐Whole-body tactile sensing[11]

    Evidence: Dense pressure, vibration and temperature sensing across the full body surface. Robot skin at whole-body scale exists in research, with many issues still open.

    Struck when: Commercial robot ships with durable whole-body tactile skin (≥90% coverage) used in closed-loop control.

Learning

  • ☐Sample-efficient motor learning[7][12]

    Evidence: RT-1: ~130k demos, 13 robots, 17 months. Open X-Embodiment: 1M+ trajectories, 22 embodiments. A cat reaches adult motor competence in weeks on one body.

    Struck when: Robot learns a novel motor skill to reliable performance in <20 real-world trials, no simulation, no teleoperated demos.

  • ☐Zero-shot generalisation[13]

    Evidence: Robot-relevant training data is scarce; safety guarantees, uncertainty quantification and real-time execution remain open for robot foundation models.

    Struck when: Robot performs open-ended tasks in a previously unseen home with no task-specific data or fine-tuning.

  • ☐Sim-to-real transfer[14]

    Evidence: Policies trained in simulation degrade on hardware. Domain randomisation, domain adaptation and meta-learning narrow the gap; none close it.

    Struck when: Policy trained purely in simulation runs on hardware with no measurable performance drop across contact-rich tasks.

Energy

  • ☐Compute energy efficiency[1][2][5]

    Evidence: Cat brain: 1.2 billion neurons on ≈0.3 W (est.). Jetson T5000 robot computer: up to 130 W. Ratio ≈1:400, and the cat's figure includes locomotion control.

    Struck when: Onboard controller runs perception, planning and whole-body control for a cat-scale robot at ≤1 W.

  • ☐Whole-body power budget[3][4]

    Evidence: Lean 4.5 kg cat: ≈274 kcal/day ≈ 13 W average, all subsystems. Spot: 564 Wh over ~90 min ≈ 380 W.

    Struck when: Cat-scale legged robot sustains a full day of cat-equivalent activity at ≤15 W average draw.

  • ☐Onboard energy density[15][16]

    Evidence: Body fat: 9 kcal/g ≈ 10,500 Wh/kg. Commercial Li-ion cells: ~270 Wh/kg; practical chemistry limit ~400 Wh/kg. Roughly a 40× gap.

    Struck when: Shipping robot carries energy storage above 2,000 Wh/kg at pack level.

  • ☐Energetic autonomy[17]

    Evidence: Best result: EcoBot-II on microbial fuel cells fed flies, 12 days of operation with one radio transmission every ~14 min.

    Struck when: Mobile robot runs 30 days at useful activity levels solely on energy foraged from its environment.

Maintenance

  • ☐Autonomous self-healing[8][18]

    Evidence: Diels-Alder polymer actuators recover after cuts when heated. In most cases human intervention is still required; the full damage → assessment → repair loop is unstudied.

    Struck when: Robot detects, assesses and repairs its own structural damage with no human intervention, then resumes duty.

  • ☐Decade-scale service life[6]

    Evidence: Cat median longevity: 14.0 years, no scheduled maintenance, continuous tissue turnover. Machines wear: Li-ion capacity fades with cycle count; gears and bearings need replacement.

    Struck when: Legged robot model reaches a 10-year median service life in daily use on original actuators and energy storage.

§4 Judgment Day log

Abilities lost to the machines

No Judgment Days on record. Humanity remains in the care of cats.

§5 Sources

References

  1. Jardim-Messeder D. et al. (2017). Dogs have the most neurons, though not the largest brain: trade-off between body mass and number of neurons in the cerebral cortex of large carnivoran species. Frontiers in Neuroanatomy 11:118. doi.org/10.3389/fnana.2017.00118
  2. Herculano-Houzel S. (2011). Scaling of brain metabolism with a fixed energy budget per neuron. PLoS ONE 6(3): e17514. doi.org/10.1371/journal.pone.0017514
  3. National Research Council (2006). Nutrient Requirements of Dogs and Cats. National Academies Press. (Maintenance energy for lean adult cats: 100 kcal × kg BW^0.67 per day.) doi.org/10.17226/10668
  4. Boston Dynamics. Spot specifications: 564 Wh battery, typical runtime ~90 minutes. support.bostondynamics.com/articles/Knowledge/Spot-Specifications-49916
  5. NVIDIA (2025). Jetson AGX Thor: Jetson T5000 module, up to 2,070 FP4 TFLOPS at a 130 W power envelope. www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
  6. O'Neill D.G. et al. (2015). Longevity and mortality of cats attending primary care veterinary practices in England. Journal of Feline Medicine and Surgery 17(2): 125–133. doi.org/10.1177/1098612X14536176
  7. Brohan A. et al. (2022). RT-1: Robotics Transformer for real-world control at scale. arXiv:2212.06817. arxiv.org/abs/2212.06817
  8. Kashef Tabrizian S., Terryn S., Vanderborght B. (2025). Toward autonomous self-healing in soft robotics: a review and perspective for future research. Advanced Intelligent Systems. pmc.ncbi.nlm.nih.gov/articles/PMC12370171/
  9. Mirvakili S.M., Hunter I.W. (2018). Artificial muscles: mechanisms, applications, and challenges. Advanced Materials 30(6): 1704407. doi.org/10.1002/adma.201704407
  10. Billard A., Kragic D. (2019). Trends and challenges in robot manipulation. Science 364(6446): eaat8414. doi.org/10.1126/science.aat8414
  11. Dahiya R.S., Metta G., Valle M., Sandini G. (2010). Tactile sensing — from humans to humanoids. IEEE Transactions on Robotics 26(1): 1–20. doi.org/10.1109/TRO.2009.2033627
  12. Open X-Embodiment Collaboration (2023). Open X-Embodiment: robotic learning datasets and RT-X models. arXiv:2310.08864. arxiv.org/abs/2310.08864
  13. Firoozi R. et al. (2024). Foundation models in robotics: applications, challenges, and the future. International Journal of Robotics Research. arxiv.org/abs/2312.07843
  14. Zhao W., Peña Queralta J., Westerlund T. (2020). Sim-to-real transfer in deep reinforcement learning for robotics: a survey. IEEE SSCI 2020: 737–744. doi.org/10.1109/SSCI47803.2020.9308468
  15. Ulvestad A. (2018). A brief review of current lithium ion battery technology and potential solid state battery technologies. arXiv:1803.04317. arxiv.org/abs/1803.04317
  16. FAO (2003). Food energy: methods of analysis and conversion factors. FAO Food and Nutrition Paper 77. (Fat: 9 kcal/g.) www.fao.org/4/y5022e/y5022e00.htm
  17. Melhuish C., Ieropoulos I., Greenman J., Horsfield I. (2006). Energetically autonomous robots: food for thought. Autonomous Robots 21: 187–198. (EcoBot-II.) doi.org/10.1007/s10514-006-6574-5
  18. Terryn S., Brancart J., Lefeber D., Van Assche G., Vanderborght B. (2017). Self-healing soft pneumatic robots. Science Robotics 2(9): eaan4268. doi.org/10.1126/scirobotics.aan4268

Method note (brain power): 1,215 million neurons in the cat brain[1] × 5.79 × 10⁻⁹ µmol glucose per neuron per minute[2] ≈ 7.0 µmol/min, × ≈2.8 J per µmol of glucose oxidised ≈ 0.33 W. The same arithmetic gives ≈22 W for a human brain, close to the commonly cited 20 W. The per-neuron figure was measured in rodents and primates, so treat the cat number as an estimate. Whole-body power: 100 kcal × 4.50.67 ≈ 274 kcal/day ≈ 13 W.[3]