Skip to content
RobotWorld
← Back

Zetta ζ: Closed-Loop Recovery Harness on Frozen VLA Policies

Loading video
Loading video

Tsinghua AIR introduces Zetta ζ, a closed-loop embodied harness that keeps base VLA policies frozen while evolving code-based runtime critics and recovery skills online — robots detect failures, recover mid-task, and reuse what worked. LIBERO-Pro success jumps 31%→90.8% with 11.1x inference speedup.

ZettaVLA失败恢复自我进化清华清华AIRLIBERO操作manipulationclosed-loopTwitter
Category: research
Author: @techniahqrobot
Date: 2026-08-30T00:00:00
Duration: 16.533s
Source: https://x.com/techniahqrobot/status/2094116103513391294
Reference: https://arxiv.org/abs/2608.16590

Related Videos

mimic-video Beats pi0 VLAs on LIBERO
0:09

mimic-video Beats pi0 VLAs on LIBERO

@cybernetic_lab · 3 views · 2026-07-29
VLAmimic-videoLIBERO
Patch Policy: pretrained ViT beats OpenVLA-OFT with 0.7% params
0:22

Patch Policy: pretrained ViT beats OpenVLA-OFT with 0.7% params

@jeffacce · 4 views · 2026-07-22
VLAPatch PolicyViT
Xiaomi Xiaomi-Robotics-1 VLA Scaling Laws
5:04

Xiaomi Xiaomi-Robotics-1 VLA Scaling Laws

@GranzenTon · 4 views · 2026-07-20
VLA缩放定律Xiaomi
RoboChemist: Fully Autonomous Wet-Lab Chemistry Robot
0:49

RoboChemist: Fully Autonomous Wet-Lab Chemistry Robot

@LeoKharon · 49 views · 2026-08-30
RoboChemist实验室自动化清华AIR
Flex-π: Multi-Stream World-Action Model
0:50

Flex-π: Multi-Stream World-Action Model

@GeYan_21 · 9 views · 2026-08-12
VLA世界模型WAM
Flex-π: compute-flexible robot foundation model
0:10

Flex-π: compute-flexible robot foundation model

@yokoshaan · 2 views · 2026-08-14
机器人基础模型VLA世界模型