SERA: self-evolving VLM harnesses tackle long-horizon robotic manipulation
Tingwen Du · hf · 2026-10-22
Researchers from USTC, Shanghai AI Lab, SJTU and others propose SERA to address long-horizon robotic manipulation, where non-Markovian dependencies require retaining task progress and past interaction outcomes while partial observability and execution errors compound over time. SERA uses a hierarchical design: an editable harness maintains task context and lets the VLM planner self-evolve its planning through execution feedback, paired with a fixed low-level VLA executor.
More from Embodied
- Scoble tours Physical AI Summit: Wayve robotaxi, Kodiak truck and stealth robot brain startup — Scobleizer · 2026-10-22
- Waymo pauses robotaxi service in 4 cities as Tesla expands to 2 more — emmanuelvivier · 2026-10-22
- Tesla delays unsupervised FSD to 2027, limited to robotaxi zones — emmanuelvivier · 2026-10-22
- AiMOGA opens preorders for Mobo family humanoid robot through Dec 31 — emmanuelvivier · 2026-10-22
- Generalist AI's GEN-1.5 puck-packing bot shows emergent physical reasoning and backtracking — chris_j_paxton · 2026-10-22
- Industrial robotics takes center stage at NVIDIA GTC 2026 Berlin — AntonObukhov1 · 2026-10-22