Track map
🌉 Sim-to-Real
Bridge the reality gap: randomize everything, calibrate the digital twin, and transfer for real.
🤖 Humanoids
GR00T, Helix, Optimus, Unitree: the race for a general-purpose body.
🧠 Hybrid Architectures
An LLM that thinks slowly, a VLA that acts quickly: the dual-system approach in practice.
🏭 Applications
Manufacturing, logistics, healthcare, and agriculture: where the ROI already makes sense—and where it doesn’t yet.
💰 Market and Opportunities
US$ 111B by 2030: value chain, players, funding, and where the jobs are.
🚀 2026-2030 frontier
Scaling laws, world models, self-improvement, and certification: what comes next.
Detailed content
🌉 Sim-to-Real: Domain Randomization and Transfer
Train on millions of simulated episodes and run on the physical robot.
What it is: the systematic divergence between what the simulator models and what the physical world does — contact friction, actuator delay, camera noise, deformable objects.
Why it matters: a policy that gets 99% in simulation can plummet to 30% on hardware. Without mapping the sources of the gap, you optimize for the wrong place.
Key concepts: reality gap, contact dynamics, sensor noise, actuation delay, physics fidelity.
What it is: aggressively randomize sim parameters so the real world becomes “just another sample” from the training distribution.
Why it matters: is the cheapest and most robust technique for zero-shot transfer; it was what enabled robots to go from simulation to real-world manipulation.
Key concepts: visual DR, dynamics DR, automatic DR (ADR), difficulty curriculum.
What it is: instead of covering everything with noise, brings the feature distribution in sim and reality closer (GANs, adversarial, embedding alignment).
Why it matters: complements DR when the visual gap is too large to randomize.
Key concepts: domain adaptation, feature alignment, RCAN, real-in-the-loop.
What it is: measure the real robot to calibrate the simulator (masses, friction, latencies) and build a faithful digital twin.
Why it matters: reduces the gap at its source; instead of covering uncertainty, you eliminate it where possible.
Key concepts: sysID, digital twin, calibration, Isaac Sim/MuJoCo tuning.
What it is: mix cheap simulation data with a few expensive real demonstrations; simulation provides coverage, real data provides final fidelity.
Why it matters: is the dominant practical recipe—95% sim + 5% real beats either alone.
Key concepts: co-training, fine-tuning, sample efficiency, real-world adaptation.
What it is: protocols that measure performance in the real world under variation, not in the sim where you trained.
Why it matters: sim-score inflates expectations; only real A/B testing counts for deployment.
Key concepts: real-world success rate, held-out, sim2real gap metric.
🤖 Humanoids: Figure, Tesla, Unitree, GR00T
The bet on a general-purpose body for a world built by humans.
What it is: the thesis that the human form unlocks the world’s stock of tools, ladders, and environments without modifying them.
Why it matters: morphological generality + foundation models = one body, infinite tasks.
Key concepts: morphological generality, brownfield deployment, dexterity.
What it is: open foundation model for humanoids with a dual system (reasoning + action) trained on real, synthetic, and neural-trajectory data.
Why it matters: NVIDIA provides the "brain + data pipeline (Isaac/Cosmos)" adopted by several OEMs.
Key concepts: GR00T N1/N1.5, dual-system, Isaac GR00T, synthetic data.
What it is: Figure’s architecture with System 2 (slow VLM, ~7-9Hz) that reasons and System 1 (fast policy, ~200Hz) that controls dexterous hands.
Why it matters: resolves the generality × frequency trade-off; pioneered full-torso control via VLA.
Key concepts: Helix, System 1/2, latent vector, whole-upper-body.
What it is: Tesla humanoid reusing end-to-end vision and FSD compute, focusing on internal manufacturing first.
Why it matters: production scale advantage and vertically integrated teleoperation data.
Key concepts: Optimus Gen 2/3, FSD reuse, industrial scale, proprietary actuators.
What it is: low-cost Chinese humanoids (G1 starting at ~US$ 16k) that democratized research on real hardware.
Why it matters: lowers the barrier to entry; became the standard platform for locomotion/manipulation papers.
Key concepts: Unitree G1/H1, cost, accessible hardware, open SDK.
What it is: coordinate gait, balance, and arms in a single controller; RL for robust locomotion on terrain.
Why it matters: manipulating while walking is what separates a demo from real-world utility.
Key concepts: whole-body control, locomotion RL, balance, loco-manipulation.
🧠 Hybrid Architectures: LLM Planner + VLA Executor
Think slowly and act quickly in the same robot.
What it is: separate high-level planning (slow, abstract) from reactive control (fast, continuous).
Why it matters: no single network is good at both long-horizon reasoning and a 200Hz control loop.
Key concepts: hierarchy, symbolic planning, reactive policy.
What it is: A slow VLM "thinks" and emits a latent vector; a fast policy consumes it to generate high-frequency actions.
Why it matters: is the convergent architecture of leading humanoids in 2026.
Key concepts: dual-system, latent conditioning, async inference.
What it is: use an LLM to break a long instruction into executable subtasks, optionally as code.
Why it matters: unlocks long-horizon tasks without training a giant monolithic policy.
Key concepts: SayCan, Code-as-Policies, task decomposition, prompting.
What it is: connect the planner’s language to skills the robot can actually execute in its current state.
Why it matters: avoids beautiful but impossible plans; the "can" in SayCan.
Key concepts: affordance, grounding, value function, feasibility.
What it is: close the loop — detect failure, update memory, and replan instead of blindly continuing.
Why it matters: robustness comes from recovery, not perfect execution.
Key concepts: closed-loop, inner monologue, replanning, memory.
What it is: chain dozens of skills for minutes while maintaining coherence and robustness.
Why it matters: is the current bottleneck—long tasks still accumulate failures.
Key concepts: long horizon, skill composition, robustness, drift.
🏭 Applications: Manufacturing, Logistics, Healthcare, Agriculture
Where general-purpose robotics is already creating value—and where it still falls short.
What it is: flexible assembly, visual inspection, and collaborative cobots with ROI measured in payback cycles.
Why it matters: is the most mature market; VLA reduces the cost of reprogramming for each SKU.
Key concepts: cobot, payback, flexibility, inspection.
What it is: separation and packaging in warehouses, the most commercialized use case for AI-powered manipulation.
Why it matters: high volume, a semi-structured environment, and clear ROI: today’s “killer app.”
Key concepts: bin picking, Covariant RFM, Symbotic, Amazon, fulfillment.
What it is: assisted surgery, tissue manipulation and care, with extremely high safety requirements.
Why it matters: high margins, but enormous regulatory and reliability barriers.
Key concepts: assisted surgery, force control, regulation, assistance.
What it is: harvest delicate fruit, map plants, and operate in a fully unstructured environment.
Why it matters: labor shortages create demand; natural variability is the challenge.
Key concepts: selective harvesting, phenotyping, outdoor, variability.
What it is: cooking, cleaning, and tidying the house — the most variable environment and the least tolerant of social errors.
Why it matters: largest potential market, but the last to close technically.
Key concepts: open-world, domestic, generalization, long tail.
What it is: safety, total cost, reliability, and integration with legacy systems.
Why it matters: technology is rarely the bottleneck; integration and trust are.
Key concepts: TCO, MTBF, safety, integration, change management.
💰 Market and Opportunities: $111B by 2030
Value chain, players, funding, and where the jobs are.
What it is: market projections (Goldman ~US$ 38B in humanoids by 2035; broad estimates of US$ 111B+ for AI robotics by 2030).
Why it matters: sizes the opportunity and grounds the hype in sources.
Key concepts: TAM, CAGR, Goldman/BCG, 2030/2035 projections.
What it is: the value layers—silicon, actuation, foundation models, integrators, and data.
Why it matters: understanding where the margin accumulates determines where it’s worth starting a business or investing.
Key concepts: value chain, layers, integrators, data moat.
What it is: the startup landscape and its billion-dollar valuations (Physical Intelligence, Figure, Skild AI, 1X, Covariant).
Why it matters: capital signals where the market is betting on the next wave.
Key concepts: Physical Intelligence, Figure, Skild, 1X, valuation.
What it is: robot-as-a-service, model licensing, and monetization of operational data.
Why it matters: RaaS reduces the customer's CapEx and aligns revenue with the value delivered.
Key concepts: RaaS, licensing, data flywheel, recurring revenue.
What it is: in-demand profiles — VLA research, robotics MLOps, integration, teleoperation data engineering.
Why it matters: guides your career path in the sector.
Key concepts: robot MLOps, data engineering, integration, research.
What it is: the risks that could delay the curve — hype, residual technical gaps, regulation, and timing.
Why it matters: reading risks is as strategic as reading opportunities.
Key concepts: hype cycle, trough of disillusionment, timing, regulation.
🚀 2026-2030 frontier: Scaling Laws, World Models, Certification
What's still uncertain—and what changes when it's resolved.
What it is: the billion-dollar question—is performance scaling predictably with data and compute, as it does for LLMs?
Why it matters: if yes, the race becomes one of data collection; if not, the architecture still matters.
Key concepts: scaling laws, data×compute, uncertainty, embodiment.
What it is: models that learn the world's dynamics to imagine futures, generate data, and plan (Genie, NVIDIA Cosmos).
Why it matters: cheap neural simulation could break the data bottleneck.
Key concepts: world model, Genie, Cosmos, neural sim, planning.
What it is: robots collecting their own data via RL at scale and autonomous data collection.
Why it matters: the flywheel that can make continuous improvement inexpensive.
Key concepts: self-improvement, RL at scale, autonomy, flywheel.
What it is: one brain, many bodies — fleets that learn together and share improvements.
Why it matters: each robot improves them all; compounding scaling advantage.
Key concepts: fleet learning, cross-embodiment, transfer, OTA.
What it is: ensure learned robots can be certified — ISO safety standards, fail-safe behavior, physical alignment.
Why it matters: without certification, there is no large-scale deployment around humans.
Key concepts: ISO 10218/15066, fail-safe, warranties, physical alignment.
What it is: projections of the impact of general-purpose robotics on labor, productivity, and economic structure.
Why it matters: career and business decisions depend on which scenario you believe.
Key concepts: 2030 scenarios, productivity, displacement, transition.