The promise of robotics has never been more compelling — a future of industrial efficiency, labor augmentation, and everyday integration. But between today’s impressive demos and tomorrow’s scaled deployments lie critical gaps that separate potential from product.
We believe robotics is at an inflection point. While foundational technologies like LLMs and vision-language models (VLMs) have made perception and planning feel like near-solved problems, the real bottlenecks lie in the physical world. Manipulating real-world objects, ensuring safety in unstructured environments, and producing reliable hardware at scale. These are the make-or-break challenges for robotics startups today.
Robotics development is progressing through distinct capability levels:
- L1 – Single Task Reliability: Reliable single-task execution in a narrow, structured scenario. Able to reliably complete one type of action (e.g., picking one type of box from a fixed location in a warehouse)
- L2 – Limited Multi-Task Competence: Limited multi-task capability, commercially deployable with human supervision. Able to perform several tasks; partial behavior expert in a single scenario; requires human backup for edge cases; low failure rate; commercially deployable with supervision.
- L3 – Scenario Expert: Full autonomy within a single environment (e.g., end-to-end warehouse operation). Able to handle all tasks within one environment autonomously without human backup (e.g., fully autonomous warehouse operation across all product types and situations).
- L4 – Cross-Scenario Generalist: Complete all tasks across different scenarios, achieving human-level versatility across verticals (the “general-purpose humanoid” vision). Still largely aspirational. general-purpose robotics remains both technically elusive and, in many cases, economically unnecessary.
- L5 – Superhuman Performance: Exceed human capabilities in speed, precision, endurance, or reliability across multiple domains.
Where We Are: L1 to L3, but L4 Remains Distant.
Today, most commercial deployments live at L1 and L2. A handful of structured verticals are inching toward L3. But moving beyond that will require breakthroughs in both physical manipulation and platform design, especially when dealing with deformable objects and contact-rich tasks.
Another key question is how wide the gap really is between capability levels, and what “good enough” means at each stage. In industrial settings, managers often require 99.5–99.9% stable success rates before full deployment, since even brief downtime can trigger costly production delays. Achieving that reliability isn’t linear: moving from 98% to 99% feels like halfway, yet reaching 99.5% is another halfway again. In contrast, service and field robots can tolerate more errors, but their threshold depends more on safety, recovery speed, and user experience.
Gap 1: Manipulation, Data, and Physical Intelligence
Unlike LLMs, physical AI must grapple with friction, deformability, weight distribution, and uncertainty. Manipulation isn’t just about identifying an object, it’s about knowing how it behaves when touched, squeezed, or dropped. That kind of learning requires massive, high-quality data, both simulated and real.
There are two schools of thought:
- Synthetic-first: Some teams report success training on 99% synthetic data.
- Hybrid-first: Others stress that real-world data is still critical, especially in high-risk tasks where failure is costly.
Both can be right. Simulation is increasingly effective, especially in structured environments like logistics. But for unstructured settings — think food prep or home robotics — the sim-to-real gap remains wide. Physics engines just can’t model the full complexity of our messy world… yet.
What’s needed?
- Scaled teleoperation programs to collect diverse, high-quality manipulation data.
- Smarter simulation platforms to improve fidelity and reduce sim-to-real loss.
- Closed-loop control systems that self-evaluate and optimize in real time.
Critically, there’s no “boss-level” task that unlocks the rest. Mastering one use case doesn’t transfer neatly to another. Folding a towel and assembling a circuit board may both require dexterity, but the physics, constraints, and stakes are completely different.
Gap 2: Hardware and the Path to Scalable Deployment
Physical limits are just as important as algorithmic ones, and harder to ignore. Robotics teams must make deliberate co-design tradeoffs based on use case:
- Laundry folding? High-DoF fingers, robust tactile sensing, and tolerance for slip and occlusion.
- Precision assembly? Focus on stiffness, repeatability, and pose accuracy. Not fingertip finesse.
These choices shape what data can be captured and what actions can be controlled. They also introduce path dependence: hardware constraints (e.g., sensor layout, actuator precision) directly impact model performance and future adaptability.
Common challenges across the stack include:
- Tactile sensors: Hard to manufacture, harder to integrate, and even harder to interpret in real time.
- Force control: Still an open research problem, especially in delicate, adaptive manipulation.
- Component defect rates: Actuators, motors, and reducers are prone to failure and expensive to source reliably.
- Economies of scale: Robotics, like EVs, faces a long climb up the quality-cost curve. Think Tesla in 2012, not 2022.
What’s needed:
- Durable, production-grade tactile and force sensors.
- Modular, efficient sensor integration frameworks.
- Smarter processing pipelines for high-frequency, multidimensional data.
Gap 3: Safety
Beyond that, as robotics enter non-industrial spaces, safety becomes paramount. Imagine a robot folding laundry next to your young child. It would be reasonable for that to worry you.
Today, it is not a solved problem, with just sensors and an e‑stop, or teleoperation. It’s a system-level discipline built on the coordination of hardware, software, data, and certified processes that insurers and regulators accept.
The founders who win here won’t just be excellent roboticists. They’ll be full-stack systems thinkers, building hardware and software in tandem, understanding data loops intimately, and knowing where to trade perfection for progress.
If you’re a founder building at the intersection of physical intelligence, hardware innovation, and real-world deployment — especially across these critical gaps of manipulation, platform design, and scaled safety — we want to meet you. Please do reach out!
