Reinforcement Learning for Robotics
Research labs and platforms applying deep reinforcement learning directly to robot skill acquisition and control, enabling robots to learn dexterous and locomotion tasks from reward signals rather than demonstrations.
CAPITAL FIGURES ARE MEDIA-EXTRACTED ESTIMATES, NOT VERIFIED FILINGS.
EXTRACTED FROM 25+ PODCASTS & VC NEWSLETTERS · MEDIA-REPORTED FIGURES, NOT VERIFIED FILINGS
Amazon FAR is becoming robot RL's central research engine
Amazon's footprint runs through the week's strongest physical-AI papers. Max Simchowitz and Guanya Shi hold dual CMU/Amazon FAR affiliations, Koushil Sreenath contributes from CMU, and Manikantan Nambi represents Amazon's internal robotics AI leadership. Pranay Thangeda (Amazon) co-leads work with UCSD and Michigan, and Nima Fazeli straddles Michigan and Amazon. This distributed lab-plus-academia model lets Amazon seed RL, whole-body control and grasping advances without acquiring a robotics startup. It also fits Amazon's appetite for large acquisitions elsewhere, such as the $7B Decart deal.
Weight-space diffusion over policy parameters reproduces RL-trained locomotion, dancing and spin-kick policies at 1.000 best-of-100 success. The generated policies deploy zero-shot on a real Unitree G1. C-GMT adds compliance through a residual adapter on a frozen motion tracker, and a distilled whole-body policy needs no privileged object state at inference. RL is shifting from a per-task training loop to a source of reusable policy assets.
Why it matters · If policies can be generated rather than retrained, the cost per new robot skill falls and Unitree-class commercial hardware becomes the default RL testbed.
Research is reframing the humanoid body as the manipulator, with arms and torso acting like fingers and palm. It builds on DexterityGen and argues that Figure AI, Boston Dynamics and Apptronik over-focus on end-effector dexterity. A diffusion controller running at 50Hz on a single RTX 5090 shows the approach is already deployable.
Why it matters · Dexterity IP may move from hand hardware to whole-body control stacks, which favours software-first teams.
Value caching addresses soft aliasing in robot RL with a theoretical guarantee. Dexterity from Supervised Play avoids task-specific demonstrations by training a world model on collected transitions. Yet generalist policies like π0.5 fall from 85.4% single-subgoal success to 0.0% at five subgoals in simulation. Real-world success at multiple subgoals falls to 1.2% on WidowX and 5.6% on Franka.
Why it matters · The post-training and long-horizon gap is the unsolved problem that RL is best placed to close, and it is where the next funding will concentrate.
π0 and π0.5 are cited as the reference design for continuous-action VLAs. New work argues discrete action parametrizations match or beat continuous ones across RoboMimic, RoboCasa and LIBERO. Hugging Face's SmolVLA appears as an open baseline, and 20VC groups Generalist, Pi and Skild as potential foundation-model concentrators.
Why it matters · Challenges to the π0 recipe could open space for open-source and challenger VLAs, with direct consequences for foundation-model concentration.