RL model nails 'last millimeter'

Physical Intelligence published a technical post on March 19, 2026 that introduces "RL Tokens"—an extra VLA output token that feeds a compact actor–critic so precise sub-skills can be adapted online without fine-tuning the full model. (pi.website) The company reports RLT yields up to 3× speedups on the most delicate stages across four manipulation tasks and enables on-board training with hundreds of updates per second to the small policy. (pi.website) Demonstrations highlighted contact-rich challenges including aligning a screwdriver with an M3 screw, zip‑tying, and ethernet-cable insertion as the target stages that RLT specifically accelerates and stabilizes. (humanoidsdaily.com) RLT is presented as a surgical complement to Physical Intelligence’s earlier RECAP workflow and π*0.6 foundation model (paper/blog published Nov 17, 2025), which the team says more than doubled throughput and roughly halved failure rates on long‑horizon tasks during on-robot fine-tuning. (pi.website) (arxiv.org) Physical Intelligence has been posting long-run demos with π*0.6—making espresso across a daytime stretch, folding 50 novel laundry items in a house, and assembling and labeling 59 factory boxes—as evidence of the RECAP-trained model’s sustained reliability. (pi.website) The group also maintains an open-source "openpi" repository with base checkpoints and tooling (the repo shows active commits and ~10.7k stars), and the company closed a reported $600M Series B in November 2025 led by CapitalG, which the press says valued the firm at about $5.6B. (github.com) (siliconangle.com)