Astra agentic policy hits 74.8% vs 51% for specialized model on Robolab-120 robot benchmark
zakkohane · x · 2026-11-02
Zak Kohane shared details of the Robolab-120 physics-based benchmark experiment (Franka Panda 7-DoF arm, Robotiq 2F-85 gripper), comparing three policies: Direct (stateless LLM), Agentic (persistent Codex session with Python and tools), and Orchestrated (hierarchical control via Cosmos3-Nano-Policy, a WAM finetuned on DROID). Result: Agentic >> Orchestrated > Direct, pointing to a handoff problem. The specialized Lakeside model scored 51% success / 65% completion; Astra with the agentic policy hit 74.8% / 83.5% — 23.8 percentage points ahead of the SOTA specialized model.
More from Embodied
- Tech Sales Veteran Reviews iPhone Duo: Worth It, But App Support Lags — thedealdirector · 2026-11-03
- Tutor Intelligence raises $75M Series B led by a16z to deploy teleop-trained robots — a16z · 2026-11-03
- Figure ships newly manufactured F.03 humanoid robots off BotQ production line — adcock_brett · 2026-11-03
- Robotics is having its GPT-2 moment, says State of AI report author Nathan Benaich — nathanbenaich · 2026-11-03
- MIT's MH-1 from 1962: one of the earliest tactile-guided robot manipulation experiments — kevin_zakka · 2026-11-02
- Amazon-led HULK framework gives Unitree G1 humanoid the strength to pull a 300 kg cart — CyberRobooo · 2026-11-02