Differences

This shows you the differences between two versions of the page.

--- start [2026/06/24 06:25]
IIS Webadmin
+++ start [2026/06/25 06:25] (current)
IIS Webadmin
@@ Line 45: / Line 45: @@
     funktionieren. Daraus ergeben sich ein realistisches Verständnis
     für ihre Möglichkeiten und Grenzen sowie einige grundlegende
-    Empfehlungen für den Umgang mit ihnen im Schulbetrieb.</blockquote></td></tr><tr valign="top"><td>2026-03-31</td><td>Samuele Tosatto gives an invited keynote <i>Where are all the intelligent robots? A quest for efficiency in reinforcement learning</i> at <a href="https://rl4aa.github.io/RL4AA26/">Reinforcement Learning For Autonomous Accelerators 2026</a>, Liverpool. <span class="actions"><a href="javascript:void(0)" onclick="showHide('newsitem6', 'Abstract')">[Abstract]</a></span><blockquote id="newsitem6Abstract" style="display:none">As artificial intelligence reshapes our digital world at a breathtaking pace,
+    Empfehlungen für den Umgang mit ihnen im Schulbetrieb.</blockquote></td></tr><tr valign="top"><td>2026-03-31</td><td>Samuele Tosatto gives an invited keynote <i>Accelerating Reinforcement Learning with Off-Policy Data: Promises, Pitfalls, and Future Directions</i> at <a href="https://rl4aa.github.io/RL4AA26/">Reinforcement Learning For Autonomous Accelerators 2026</a>, Liverpool. <span class="actions"><a href="javascript:void(0)" onclick="showHide('newsitem6', 'Abstract')">[Abstract]</a></span><blockquote id="newsitem6Abstract" style="display:none">Reinforcement learning is a promising technique for solving complex control problems in real-world physical systems, such as robotics, plasma stabilization, and particle accelerators. However, RL is often data-hungry, and its classic on-policy formulation is often inefficient, as it disallows data reuse, and unsafe, as it requires the agent to interact with the environment from scratch.
-    a curious question arises: Where are all the real-world, intelligent robots?
+Off-policy reinforcement learning offers a more appealing paradigm by enabling the reuse of historical data and the utilization of safe, external behavior sources (such as human operator logs). However, this flexibility comes at a cost: off-policy learning introduces significant theoretical instabilities. In this talk, we will analyze some fundamental difficulties in off-policy reinforcement learning, both in value and policy learning, explore the algorithmic landscape that tames them, and see the future direction in which the field is moving. </blockquote></td></tr><tr valign="top"><td>2025-11-19</td><td>Justus Piater gives an invited talk <i>Structural Understanding – The Grand Challenge of Robot
-    While we have mastered the generation of text, images, and video by leveraging vast web-scale datasets,
-    robotics still faces a fundamental data bottleneck. Reinforcement learning (RL) offers a compelling solution,
-    enabling agents to learn autonomously by collecting their own experience.
-    However, the path to autonomy is often blocked by the staggering inefficiency of current RL algorithms,
-    which can require millions of trials to master simple tasks. This talk embarks on a quest to tackle this
-    efficiency problem head-on. I will argue that a crucial step toward unlocking the potential of
-    robot learning lies in a two-pronged approach: first, by developing statistically efficient
-    algorithms that can reuse data by leveraging sound off-policy techniques, and second,
-    by designing better action representations for physical, real-world agents.
-    By making our algorithms more efficient and refining their core hypotheses,
-    we can accelerate the journey toward real embodied artificial intelligence.</blockquote></td></tr><tr valign="top"><td>2025-11-19</td><td>Justus Piater gives an invited talk <i>Structural Understanding – The Grand Challenge of Robot
     Learning</i> at <a href="https://elliit.se/news-and-events/focus-period-lund-2025/symposium/">ELLIIT Focus Period Symposium: Robot Learning</a>, Lund University. (The ELLIIT Focus Period Symposium is the highlight of the five-week focus period, during which young international scholars, ELLIIT researchers and other well-established international academics gather in Lund to work together on joint research challenges.) <span class="actions"><a href="javascript:void(0)" onclick="showHide('newsitem7', 'Abstract')">[Abstract]</a></span><blockquote id="newsitem7Abstract" style="display:none">AI has made great progress in recent years, and the
     sophistication of robots has been rising with costs falling. Yet,

User Tools

Site Tools

Differences

Page Tools