SJTU: VLA skill recombination fails at the "start state gap"
AutoLab-SJTU · hf · 2026-10-15
SJTU researchers probe combinatorial generalization limits in embodied VLA models: full-instruction training reaches 40.0% ordered success on covered combinations but only 7.7% on recombinations — memorizing sequence templates rather than truly recombining skills.
Contributions:
- HSE: a diagnostic framework decomposing long-horizon failures across planner, trigger, and policy
- RoboCombine benchmark: failures concentrate at subtask transitions where downstream tasks launch from upstream handoff states instead of canonical starts — the start state gap
- PRTI, a lightweight atomic-task post-training strategy broadening start-state coverage, lifts ordered success from 4.21% (random perturbation) to 18.16% on RoboCombine
Takeaway: combinatorial generalization needs handoff-robust atomic skills, not just decomposition.
More from Embodied
- RoboParty launches RP1, a fully open-source full-stack humanoid robotics platform — GuanyaShi · 2026-10-15
- HumanoidToolBench benchmarks humanoid tool use from selection to mobile execution on real Unitree G1 — snupilab · 2026-10-15
- NYU's Temporal Flow Policy adds physical time to robot memory, beating RNN and state-space memory designs — newyorkuniversity · 2026-10-15
- G³VLA injects calibrated multi-camera geometry into VLA models to improve manipulation — Yue Peng · 2026-10-15
- Survey: Towards Agentic Robots organizes the field into 17 categories across four domains — Chonghan Qin · 2026-10-15
- Tencent ARC's Player-1 trains game agents on synthetic experience, hitting 38.3% on MineDojo-300 vs 12.6% baseline — TencentARC · 2026-10-15