To get updates from our aUToronto self-driving car team, please visit this page.
Dynamic Systems Lab News
Only 10 days to go until #IROS2023 in Detroit! Here is where you can find us
Sun. Oct. 1 at 14:00: Prof. Angela Schoellig's talk "Safe Learning in Robotics: From Learning-based Control to Safe Reinforcement Learning" at the workshop on Formal methods techniques in robotics systems: design and control. Room 140E.
Mon. Oct. 2 at 09:24: Sepehr Samavi will present "Does unpredictability influence driving behavior" in Imitation Learning, Room 330B. Paper: lnkd.in/dFyVdutm
Tue. Oct 3 at 08:42: Wenda Zhao will present "Uncertainty-Aware Gaussian Mixture Model for UWB Time Difference of Arrival Localization in Cluttered Environments" in Localization III, Room 320. Paper: lnkd.in/d2FsYpWg
... See MoreSee Less
- Likes: 5
- Shares: 1
- Comments: 0
We analyzed 6 years of data across machine learning, robotics, and control. The verdict: high-impact research is linked to open-source success. Read our latest work on reproducibility arxiv.org/abs/2308.10008 and submit to our associated workshop at CDC2023 tiny.cc/rnw9vz
Angela Schoellig Technical University of Munich University of Toronto Engineering
... See MoreSee Less
FACT: Only 2.6% of papers at CDC 2021 included open-source code 😱
We are organizing a workshop on “Benchmarking, Reproducibility, and Open-Source Code in Controls,” to be held at the IEEE Conference on Decision and Control (CDC) 2023 on December 12, 2023. We are calling for short 300-word abstract submissions on opinions/experience in improving accessibility and reproducibility of research in control through benchmarks, open source code, software/hardware platforms, and pertinent educational content.
Topics of interest include but are not limited to:
– Any open-source implementation of control algorithms
– Any benchmarks and comparisons of control approaches, or competitions
– Tools and software/hardware platforms that enable accessible and reproducible research
– Tutorials/lectures on reproducible research best practices and standards
– Educational resources to make control theory accessible
By bringing together researchers in these topics, we aim to improve the accessibility, reproducibility, comparability, usability, and visibility of research in control theory. We look forward to your contributions!
Submission link: tiny.cc/cdc23-ws-abstract
Deadline: September 18, 23:59 PST
Format: Accepted abstracts will be presented as lightning talks at the workshop
Angela Schoellig Technical University of Munich University of Toronto Engineering
... See MoreSee Less
Simultaneous localization and mapping (SLAM) with online scene change detection is very important for long-term robot deployment, as failing to detect such changes can lead to localization drifts, inaccurate maps and lost robots. However, current approaches focus on high dynamics that can be observed in consecutive frames, failing to handle long-term, incremental changes.
In our recent works accepted at #RSS2023, we tackle this challenge in semi-static environments, introducing POV-SLAM. We show that our probabilistic, object-aware method produces more stable robot pose estimates and more consistent scene reconstructions despite the scene changing significantly over time. We summarize our contributions as follows:
We propose a bimodal measurement likelihood for potentially-changing objects and derive its evidence lower bound (ELBO) for efficient inference
We design an online, object-aware SLAM framework using an Expectation-Maximization algorithm and the proposed measurement model to track incremental scene changes
We release a real-world long-term SLAM dataset captured in a warehouse over a four-month span
Thank you to Clearpath Inc, Natural Science and Engineering Research Council of Canada (NSERC), and Vector Institute for supporting our research. We look forward to seeing everyone at #RSS2023!
Check out our paper and dataset for more details.
Paper: arxiv.org/abs/2307.00488
Dataset: github.com/Viky397/TorWICDataset#The-Toronto-Warehouse-Incremental-Change-SLAM-Dataset
... See MoreSee Less
We are #hiring! Suchen Sie eine abwechslungsreiche Tätigkeit als Teamassistenz (mit Fokus Finanzen) in einem enthusiastischen und internationalen Team im Bereich der #robotik und #kuenstlicheintelligenz am Innenstadt-Campus der Technical University of Munich (Maxvorstadt, Theresienstr. 90)? Dann bewerben Sie sich bei uns am Lehrstuhl “Sicherheit, Performanz und Zuverlässigkeit für lernende Systeme” lnkd.in/eRbzsDYZ.
Hier finden Sie den LinkedIn-Job lnkd.in/e67RfH6K und die TUM Stellenausschreibung: lnkd.in/e3gAP6eK .
Wir freuen uns sehr auf Ihre Bewerbung! Fragen zu dieser Stelle beantwortet Prof. Schoellig gerne auf LinkedIn oder unter ihrer TUM-E-Mail-Adresse.
... See MoreSee Less
Explore AMSwarm: an online and scalable planner with safety behaviors for Quadrotor Swarms in Cluttered Environments.
Join us at the Poster Session of ICRA 2023, located at Poster Hall (Entrance N11), on 30 May 2023, from 08:30 - 10:10 (Multi-Robot Systems I Pod 31-33).
... See MoreSee Less

Looking forward to presenting our recent work at ICRA 2023, London. We present a novel UWB sensor placement algorithm that balance the effects of anchor-tag geometry and the NLOS measurement biases. RA-L paper on arXiv: arxiv.org/abs/2204.04508 ... See MoreSee Less
We will be presenting a poster on robotic planar pushing using force feedback at the Embracing Contacts workshop at #ICRA2023! Join us on Friday, June 2.
Paper: arxiv.org/abs/2305.11048
Workshop: sites.google.com/view/icra2023embracingcontacts/home
Angela Schoellig University of Toronto Engineering Technical University of Munich
... See MoreSee Less
We are excited to announce our lab's third accepted #ICRA2023 paper by Alan Li: Multi-View Keypoints for Reliable 6D Object Pose Estimation
In this work, we propose a novel multi-view approach to object pose estimation, leveraging known camera transformations from an eye-in-hand setup to combine heatmap and keypoint estimates into a probability density map over 3D space.
Read more here: lnkd.in/g66WfJFV
... See MoreSee Less
Distributed approaches based on Sequential Convex Programming (SCP) have been the state-of-the-art for quadrotor swarm motion planning. However, these approaches struggle in cluttered environments due to conservative approximations made in the optimization problem, namely linearization of collision constraints and axis-wise decoupling of kinematic bounds. Such approximations are made to obtain a quadratic program (QP) at the expense of small feasible sets.
We are thrilled to announce our lab's second accepted #ICRA2023 paper! This work, led by Vivek Adajania, solves the optimization problem as a QP without relying on previously mentioned approximations, achieving superior success rate, mission time, and per-agent computation time.
arXiv: arxiv.org/pdf/2303.04856.pdf
Video: youtu.be/eIBcOKq5_Jk
... See MoreSee Less