5/810 → 16/622; +1.96 percentage points of formation share.
Underneath the shared "reinforcement learning" label, this batch splits into two distinct businesses: companies selling the picks and shovels of post-training, like HUD and idler building the environments and evals that frontier labs consume, and companies selling the finished mine, like Osmosis, Freesolo, and Sylvian, which fine-tune and deploy specialized agents directly for enterprise customers. That split explains why Data Labeling, growing at 2.60x, trails this formation's 4.17x rather than tracking it in lockstep; the market has moved past labeling static data and into generating dynamic training signal, which is a more valuable and more defensible position. Infrastructure, growing at a comparatively sedate 1.48x, lags even further behind, suggesting the bottleneck has shifted away from raw compute or tooling and toward the harder problem of producing good reward signal and realistic environments, the exact work Cartpole and Aviro describe doing across thousands of tool steps. That shift also surfaces a real tension: companies like Cortex AI are pushing RL into embodied, physical-world data, yet the open question of whether policies trained heavily in simulation actually survive sensor noise, wear, and rare real conditions remains unresolved, meaning the field is scaling training capacity faster than it is scaling proof that the training generalizes. Much of this became viable only after reasoning-focused RL techniques proved out publicly in 2024 and 2025, giving startups a template for applying RL post-training to narrow domains rather than only to frontier general models. The result is a formation less about training one better chatbot and more about building the specialized data and environment supply chain that every domain-specific agent now needs.
More simulation does not by itself show that policies survive sensor noise, environmental variation, hardware wear, or rare real-world conditions.
Compare simulator, robotics, and deployment workflows for independent benchmarks, transfer metrics, scenario coverage, and reproducible failure suites.
Companies like HUD, Osmosis, and Cartpole are building infrastructure and data-labeling layers specifically for RL, suggesting demand for standardized, shareable training environments rather than each startup building bespoke simulators (as Topological is doing for robotics/3D printing).
Check back around May 2027 for new YC batches, to see whether independent evidence for "RL environment/evaluation marketplaces for agent training" has appeared.
Topological's combination of Reinforcement Learning, Robotics, and 3D Printing tags signals a bridge between simulated RL training and physical hardware deployment, which typically requires specialized sim-to-real calibration tools not yet represented among the other B2B/software-only RL companies.
Check back around August 2027 for new YC batches, to see whether independent evidence for "Robotics-specific RL simulation-to-real transfer tooling" has appeared.
Nuntius and idler pair RL with Infrastructure/Engineering tags, indicating teams need visibility into training runs and agent behavior in production; this mirrors how MLOps observability emerged after general ML infra matured.
Check back around April 2027 for new YC batches, to see whether independent evidence for "RL-specific observability/monitoring dashboards" has appeared.
With TrainLoop, Agentin AI, and Kairos focused on enterprise RL pipelines, the proliferation of custom reward functions across B2B deployments creates a need for third-party tools that verify reward models aren't exploited or misaligned before production rollout.
Check back around June 2027 for new YC batches, to see whether independent evidence for "RL reward-model auditing and verification tools" has appeared.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
Included in YC's Reinforcement Learning membership during the measured recent window.
A primary internal company segment within Reinforcement Learning, derived deterministically from the shared yc industry “Engineering, Product and Design”. This is a classification of current window members, not a causal layer or a separate time trend.
Measured-window members of Reinforcement Learning that could not be assigned to a narrower shared YC tag or industry without forcing a classification.
A primary internal company segment within Reinforcement Learning, derived deterministically from the shared yc tag “Infrastructure”. This is a classification of current window members, not a causal layer or a separate time trend.
These share a word with “Reinforcement Learning” but are tracked as separate categories. The real overlap below is the actual share of Reinforcement Learning’s companies also found in that category — not a guess.
Cambrian Signal, “Reinforcement Learning,” publication yc-2026-10-02-semantic-v8, snapshot October 2, 2026.
cambrian.jafar.com/formation/reinforcement-learning