VERA paper undercuts its own rhetoric: structured Jacobian-based IDM beats raw 14B video model for robot control
GeorgiaChal · x · 2026-09-08
Pushing back on the claim that "intermediate representations/structure no longer matter," GeorgiaChal cites Sitzmann's own VERA paper as a counterexample:
- A 14B video model is not sufficient to control the robot
- VERA introduces an embodiment-specific inverse-dynamics model (IDM) built on the robot Jacobian
- The paper explicitly shows this structured IDM has a better data–accuracy frontier than unstructured alternatives
The takeaway: at least for robot control, explicit structure still delivers measurable gains — quite far from the rhetoric.
More from Embodied
- RL-X adds BeyondMimic-style G1 motion tracking in MuJoCo and MJX Warp — Jan_R_Peters · 2026-09-08
- XPeng's Turing chip strategy: 3,000 TOPS Robotaxi, 2,250 TOPS humanoid robot — bookwormengr · 2026-09-08
- XPeng bets on end-to-end VLA 2.0 to power its GX Robotaxi commercial rollout — bookwormengr · 2026-09-08
- Ex-DeepMind's Danijar Hafner builds stealth startup teaching humanoids to plan in unseen rooms — nordicinst · 2026-09-08
- Festo's Bionics series: flapping-wing robots and swarm-intelligence bionic ants — lukas_m_ziegler · 2026-09-08
- Phillip Isola: Computer-Use Agents Paving the Way for 'Robot-Use Agents' — CSProfKGD · 2026-09-08