LIBERO-MAX: Robot Policies Drop 11–26 Points When the World Changes Mid-Task
drmapavone · x · 2026-10-27
Researchers from Tulane, NYU, Stanford, CMU and others introduce LIBERO-MAX, a benchmark testing whether robot policies can adapt when the world changes mid-execution.
- Evaluates 14 state-of-the-art policies across 8,000 cases, injecting a controlled perturbation after the robot has started acting
- Perturbation types: geometry (target/receptacle relocation), observation (camera shift, sensor noise), appearance & clutter (lighting, distractors), path constraints (obstacle insertion)
- Task success drops by 11–26 percentage points, worst under geometry and observation changes; simply updating policies more often doesn't fix it
Takeaway: adapting to a changing world remains a key unsolved challenge for deploying Physical AI. Project page: liberomax.github.io
More from Embodied
- Re-uploading a 239GB robot dataset moved only 2.6MB — inside a petabyte-scale data pipeline — mervenoyann · 2026-10-27
- ROSCon UK photo tour: BBC's whale robot, da Vinci surgical robot and the OG PR2 — kamathsblog · 2026-10-27
- Teardown of a humanoid robot Z-axis linear actuator, with Zmotor Q&A — MarwaEldiwiny · 2026-10-27
- GPT-6 Astra Beats Gemini 3.8 Flash in Paired Robot Test, Yet Neither Completes the Task — qinzytech · 2026-10-27
- Teardown of Zmotor LA500 humanoid linear actuator: 324 g, 500 N peak thrust — MarwaEldiwiny · 2026-10-27
- Agility's Digit robot to expand into two more GXO US warehouses from early 2027 — jonstephens85 · 2026-10-27