PDRVLA
VLA from Peking University and JD.com that adds a Physical Dynamics Module, which learns physical-dynamics representations, to a Qwen3.5-4B backbone without generating video at inference. It is evaluated in simulation only, and no code or weights are linked.
Peking University & JD.com · Oct 27, 2026 · VLAReport an error
99.3%LIBERO · #6 of 83
90.7%LIBERO+ · #4 of 37
81.5%Simpler WX · #6 of 32
4Benchmarks reported
Benchmark results4
LIBERO-PlusSimulation
- Paper / model card ↗Oct 2026
#4 of 3790.7%, self-reported by the authors, not a leaderboard entry
- 93.6%, self-reported by the authors, not a leaderboard entryPaper / model card ↗Oct 2026· self-reported, 50 clean + 500 randomized demos· n = 10000· non-standard
93.6%, self-reported by the authors, not a leaderboard entrynot ranked
- Paper / model card ↗Oct 2026· n = 2000
#6 of 8399.3%, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Oct 2026· visual matching
#6 of 3281.5%, self-reported by the authors, not a leaderboard entry
Not reported on 11 benchmarks
Robustness drop: 99.3% on LIBERO → 90.7% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.
Wrong or missing number? Report an error