First systematic benchmark: 9 frontier LLMs controlling real robot arms, 54-day doubling
pwlot · x · 2026-10-25
The arXiv paper "The Bitter Lesson Comes to Robotics" (2610.25287, submitted to the TAG workshop at CoRL 2026) presents the first systematic comparison of frontier multimodal LMs directly controlling physical robotic arms.
- 9 models evaluated at multiple reasoning-effort and service-tier settings on 6 manipulation tasks from the RoboDojo benchmark, on real hardware
- Current models can perform varied manipulation tasks but lack the precision and efficiency to finish within a practical action budget
- Progress is rapid: July–October 2026 performance doubling time of 53.9 days; best execution rate up 4.4x, lowest execution cost down 6.2x
- Authors argue the Bitter Lesson is coming to robotics: general-purpose LMs may one day match specialized models
Related event: Bitter Lesson Hits Robotics: LLM Arm Control Doubles in 54 Days(3 posts)→
More from Embodied
- AGI House stocks up on NVIDIA Jetson AGX Thor developer kits for upcoming build event — agihouse_org · 2026-10-25
- StackChan desk robot runs Nous Research hermes-agent fully locally with snappy voice — alexcovo_eth · 2026-10-25
- AppetiteCheck study: vagus stimulation cuts food intake 9.6% and eating speed 23.6% — plopesresearch · 2026-10-25
- EPFL's LogSync: a 4x2mm brain implant chip decoding cognitive states under 200 microwatts — petrusenko_max · 2026-10-25
- Samsung foldables at 3% after 7 years; Apple foldable predicted to hit 20% by 2028 — mitchdeg · 2026-10-25
- Figure CEO shows off three generations of its humanoid robots — adcock_brett · 2026-10-25