Inria is France's national public research institute dedicated to digital science and technology, where over 3,800 scientists work across areas including artificial intelligence, cybersecurity, mathematics, and quantum computing. It advances foundational and applied research, publishes open-source software, spins out deeptech startups, and coordinates France's national AI research strategy in partnership with universities and industry.
“The Logit-Normal (LN) noise schedule is a drop-in replacement for the Beta schedule used in π0, π0.5, and SmolVLA that reallocates 6× more gradient signal to the contact-correction regime (τ < 0.2)”
Source→“Referenced in Section 4.1 alongside π0 and π0.5 as using the problematic Beta schedule”
Source→“π₀ scores 55% on RLBench-10Tasks versus PointACT's 82.3% (Table VIII). On the real SO-100 robot, π₀ achieves 2/10 on the 'Put Sock in Drawer' task versus PointACT's 9/10.”
Source→“We utilize Qwen2.5-VL as the VLM backbone (Section IV-A). The frozen Qwen2.5-VL provides the semantic understanding layer while PointACT's action expert handles geometry.”
Source→“PointACT's architectural insight is clean: don't touch the pretrained language-vision backbone. Instead, build a separate 'point-action expert' that lets 3D geometry directly condition the action generation process at multiple spatial scales.”
Source→“Senior author, recipient of the Körber European Science Prize (acknowledged in paper). One of the most cited computer vision researchers in Europe.”
Source→“The paper runs on two real hardware platforms — a low-cost SO-100 arm and an industrial UR5 — and the results hold up.”
Source→“We initialize the encoder from a PTv3 model pretrained via large-scale self-supervised learning on building-level 3D scenes (Section III-B).”
Source→“PointACT achieves 82.3% on RLBench-10Tasks versus 73.2% for EO1 (a strong 3B parameter 2D VLA), while running at 6.7 Hz versus EO1's 3.8 Hz (Table VIII, Appendix A).”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.