Theses: Robotics & Humanoid Automation

Evolving beliefs with evidence. Confidence changes over time as new research arrives.

Thesis 1: Humanoid robots will be deployed in 50% of large factories by 2030

Boston Dynamics and Tesla have moved from demos to production commitments. Tesla's self-deployment model (robots building robots) could create an exponential scaling flywheel. The convergence of 99% sim-to-real transfer with foundation model control eliminates the two hardest barriers.

Confidence: 5/10 Supporting evidence:

• Boston Dynamics electric Atlas: 56 DOF, 50kg lift, $150K, 30K/year factory planned for 2028 Evidence: strong (Atlas)
• Tesla: 1,000+ Optimus Gen 3 deployed, 50-100K target 2026, 10M/year factory under construction Evidence: strong (Optimus)
• 99% sim-to-real correlation at production scale (ABB + NVIDIA) Evidence: strong (HyperReality)
• Foundation models as reasoning layer enables zero-shot task adaptation Evidence: strong (Humanoid-COA)

Challenging evidence:

• 2-3 year ROI payback at $150K is not yet demonstrated for enterprise customers
• 24/7 factory reliability is unproven — current deployments are supervised
• Long-horizon combined tasks still only 56-63% success rate
• Workforce displacement regulatory responses could slow adoption
• "50% of large factories" is extremely aggressive — factories retool slowly

Evolution:

• Apr 5, 2026 — Initial thesis at 5/10. The production commitments are real but "50% of large factories" by 2030 requires both technology reliability AND enterprise purchasing cycles to align. Tesla's self-deployment is the wildcard — if robots-building-robots works, scaling could be nonlinear.

Depends on: humanoid-loco-manipulation, sim-to-real-transfer, foundation-models-for-robotics Would change if: Tesla's Optimus achieves demonstrated 24/7 reliability in its own factories, or if early enterprise customers report sub-2-year payback.

---

Thesis 2: Consumer humanoids ($20K-$50K) will achieve meaningful household utility by 2028

1X NEO ($20K, Q2 2026) and Figure AI Helix 02 (household tasks from mocap) represent the first serious consumer attempts. World models + teleoperation data pipelines + progressive autonomy is a viable scaling strategy.

Confidence: 4/10 Supporting evidence:

• 1X NEO preorders at $20K consumer price point, Q2 2026 delivery Evidence: moderate (NEO)
• Figure Helix 02 handles dishwasher, laundry from mocap-trained models Evidence: moderate (Figure)
• White House demo signals political legitimacy for consumer robotics Evidence: weak (Figure)
• World model enables self-teaching through observation Evidence: moderate (NEO)

Challenging evidence:

• Safety in unstructured home environments is largely unsolved
• $20K is aspirational — $499/mo subscription may be more realistic near-term
• Task generalization beyond demonstrated capabilities remains narrow
• Home environments are far more varied than factories — each home is unique
• "Meaningful household utility" is a high bar — must do more than one or two tasks

Evolution:

• Apr 5, 2026 — Initial thesis at 4/10. The price points and demos are encouraging but homes are much harder than factories. "Meaningful utility" by 2028 probably means 5-10 reliable household tasks, which requires both hardware reliability and foundation model generalization. Most likely outcome: narrow utility (laundry, dishes) rather than general household help.

Depends on: imitation-learning, world-models, whole-body-control Would change if: 1X NEO early adopter reviews show reliable daily use for 3+ household tasks, or if safety incidents in homes cause regulatory backlash.

---

Thesis 3: Foundation models (VLMs) will replace task-specific training as the dominant control paradigm within 3 years

Humanoid-COA's 96.6% grasping and 90% mobile pick via GPT-4V — with zero task-specific training — is the strongest evidence that the future of robot control is general-purpose reasoning, not per-task RL policies.

Confidence: 7/10 Supporting evidence:

• 96.6% zero-shot grasping on physical Unitree robots via foundation models Evidence: strong (Humanoid-COA)
• LLMs autonomously designing reward functions for tactile manipulation (Text2Touch) Evidence: strong (Text2Touch)
• Pattern: VLM reasoning -> task decomposition -> pre-trained execution becoming standard Evidence: strong (Frontier)
• Teleoperation data + simulation scaling works for bootstrapping Evidence: moderate (1X NEO, Figure)

Challenging evidence:

• Long-horizon combined tasks still 56-63% success — gap is significant for production use
• Dependence on external APIs introduces latency and availability risks
• Contact-rich manipulation (assembly, cooking) may still require specialized policies
• LLM-designed reward functions are untested for fine force control tasks

Evolution:

• Apr 5, 2026 — Initial thesis at 7/10. The 96.6% grasping result is remarkable for zero-shot. The remaining gap is in complex, multi-step, contact-rich tasks — but the trajectory is clear. "Dominant paradigm" means >50% of new deployments use foundation models as the primary reasoning layer.

Depends on: foundation-models-for-robotics, humanoid-loco-manipulation, tactile-sensing Would change if: Long-horizon task success rates stagnate below 80%, or if a non-VLM approach (e.g., pure RL + world models) achieves superior performance.