Robot success rate doubles 23%→51% from pretraining data alone, no fine-tuning
rohanpaul_ai · x · 2026-11-06
Verne's V0.5 robot manipulation model shows striking data scaling: feeding it more pretraining data with zero task-specific fine-tuning roughly doubled RoboTwin 2.0 success from 23% to 51%. Per the paper's Figure 3, checkpoints evaluated on 1,000 episodes show success climbing near-linearly with data with no sign of flattening — pointing to a data scaling law for robot manipulation. That's why Verne is now scaling its data 10x.
More from Embodied
- GPT-6 Astra Tops Robot Leaderboard but Went 0/11 on Real Plug Insertions — AIFlow_ML · 2026-11-06
- Hugging Face's RoboColosseum: 5,699 blind trials show AI agents rival VLA policies, no humanoid task completed — DJiafei · 2026-11-06
- RoboColosseum benchmark: 100+ robots in 140+ real scenes, all rollouts open-sourced, 10K free eval hours for academia — DJiafei · 2026-11-06
- moonlake previews research on scaling robot learning with simulation in the RL era — du_yilun · 2026-11-06
- Nearly 100 testers praise Sphenomenon MR demo that turns a whole house into voxels — Scobleizer · 2026-11-06
- Spectral Skills lands on real robots: composing motion skills via energy functions — daibond_alpha · 2026-11-06