Reinforcement Learning

RL's center of gravity is robot control and embodied AI

Most RL mentions in the last weeks come from physical-AI research rather than language models. Papers cover compliant whole-body humanoid control (C-GMT from MIT's Improbable AI Lab), time-aware robot policies (Chronos Policy), and discrete-versus-continuous action heads for VLAs, including OpenVLA, Octo and RT-1/RT-2 lineages. Carnegie Mellon (Sreenath, Shi, Simchowitz), Stanford and MIT labs supply much of the work. Commercial humanoid players such as Apptronik and 1X appear as reference points, not as RL funding recipients.

Gemini Robotics, RT-1/RT-2, PaliGemma, Genie 2 and the new Genie 4 world model, and MuJoCo Warp (co-developed with NVIDIA) all appear as baselines or infrastructure that other research builds on. Gemini models are the default planner or verifier in many systems. DeepMind's simulation infrastructure underpins the parallel training used across the field.

Why it matters · Startups in RL and robotics are building on, or competing against, DeepMind's open and closed components, which shapes defensibility and buyer dependency.

An $35M seed with Lightspeed, Sequoia, Index, Accel and Eric Schmidt shows top-tier funds backing a new lab. ReflectionAI's founders came from the RL team behind Gemini 1 and 1.5, and AlphaGo and AlphaZero veterans are cited as the template for RL scaling. Silver's AlphaGo lineage and Assael's backing show how the network recycles talent into new companies.

Why it matters · Pedigree is the main underwriting signal in RL, so deal access depends on proximity to the DeepMind alumni network.

AlphaZero beating human-in-the-loop AlphaGo within hours is cited as the canonical case. Sakana AI and UC Berkeley's MASS lets a Qwen3.6-27B model act as solver, optimizer and judge and fine-tune on its own best runs. Commentators argue this could shift the economics of who can build capable systems.

Why it matters · If small open models can self-improve, RL environment and data vendors face pricing pressure, but compute-efficient post-training tooling gains value.

Weekly capital was $117M in the week of Aug 24, $35M in the week of Oct 5, and $72M in the week of Oct 12, with nothing since. The Ascensia Series B of $36M accounts for much of the October capital because it was reported by both StrictlyVC and PitchBook. Several rounds list no amount, including a Jaan Tallinn seed and a Samsung grant, and the 90-day stage mix is seed- and Series A-heavy.

Why it matters · Few priced RL rounds means valuations are hard to benchmark, and early-stage investors set terms with little market data.