Dyna Robotics shows RL vs IK head-to-head on wheeled robot, unveils workflow agent Dyna-2.1
chris_j_paxton · x · 2026-11-10
Dyna Robotics released a direct comparison of IK vs RL tracking the exact same ground-truth motion on their wheeled robot, responding to skepticism about whether RL is really used:
- Lower overall motion tracking error with RL
- More natural, human-like whole-body motion vs. relatively rigid IK
- Better handling of unusual end-effector poses, where RL finds better whole-body configurations
The author acknowledges why skeptics exist: many systems treat loco-manipulation as locomotion + manipulation with separate controllers and limited waist/lower-body involvement — in that architecture, whole-body RL indeed adds little.
The post also introduces Dyna-2.1, a physical agent built on Taku, a semi-humanoid robot with a learning system that leverages human experience for teachability. Key thesis: customers pay for a role, not a task — stationary tasks like 24-hour napkin folding still need a human babysitter to refill bins, while end-to-end workflows allow async handoff with minimal oversight. They released a one-hour uncut demo of an automated end-to-end workflow.
More from Embodied
- Claude Fable 5.2 scores 64.9% on RoboDojo robot control, dominating rivals — burny_tech · 2026-11-10
- AI researcher Chris Paxton comes away positive from a Tesla Cybercab ride — chris_j_paxton · 2026-11-10
- Two CoRL papers speed up VLAs 2.4x and 4.7x, fast enough for real-time robot control — songhan_mit · 2026-11-10
- ESP-WiFi-Jam-Detector: standalone 2.4/5 GHz interference detector on ESP32-C5 — Kyrannio · 2026-11-10
- Lamport Robotics open-sources 12 audited SO-100/101 robot manipulation datasets — R4nsack · 2026-11-10
- Robot dexterous hands showcased at CoRL — YuXiang_IRVL · 2026-11-10