Tsinghua's ParallelWorld uses test-time scaling for embodied agents, lifting success 56% to 83%
jiqizhixin · x · 2026-11-07
Researchers from Tsinghua University and NUS introduce ParallelWorld, a prospective planning framework for agentic embodied reasoning in a revised preprint.
Instead of executing the first plausible high-level action, the planner proposes alternative futures: a simulator or world model produces candidate observations, and an agentic verifier evaluates information gain, ranks branches, prunes weak ones, and guides further expansion. Only the first action of the selected trajectory is executed, then the system observes the real outcome and replans.
On LIBERO-Pro's Spatial position-swap setting, it raises success from HarnessVLA's 56% to 83% over 100 episodes — though the gain is specific to one perturbation setting, and a search-width ablation covers only three tasks and thirty episodes.
More from Embodied
- Xiaomi humanoids sort car parts for eight hours in factory endurance test — emmanuelvivier · 2026-11-07
- Robot hand deals cards like a Vegas dealer, researchers impressed — micoolcho · 2026-11-07
- Robotic arms are precise but still fumble soft, unpredictable objects — TansuYegen · 2026-11-07
- Realbotix Echo robots cost $11,950-$59,950: selling a humanlike interface, bodies as upgrades — TansuYegen · 2026-11-07
- Autonomous factory trains robots in 4 weeks to bend wires and screws, lands 3-year production contract — Scobleizer · 2026-11-07
- 3D tactile imaging demo at CoRL: Technion team shows artificial touch work — AvivTamar1 · 2026-11-07