Reliable Autonomous Systems Lab at MIT (REALM)
@mit_REALM
Followers
163
Following
18
Media
9
Statuses
29
Reliable Autonomous Systems Lab (REALM) @MIT. We design, analyze and verify safe control systems.
Cambridge, Massachusetts
Joined September 2022
Our code is open source! Code: https://t.co/h85OKkETRW Arxiv: https://t.co/WVqStePOqI Web: https://t.co/ZUGKu0DTFY Video: https://t.co/Ue9sh6owW8 🧵 (7/7)
0
0
1
By predicting a diverse set of possible failures prior to deployment, we can make sure robots of all sizes (from drones to power grids) stay safe. 🧵 (6/7)
1
0
1
Finding safety critical bugs in autonomous systems 🤖😈 Our next paper https://t.co/ZUGKu0DTFY aims to find --- and fix --- bugs in autonomous systems before they become a problem. 🧵 (5/7)
1
0
1
Our code for GCBF is open source! Code: https://t.co/hgYcWP34Yr Web: https://t.co/ef08z4jfN6 🧵 (4/7)
1
0
1
Trained with only 16 agents, GCBF can achieve up to 3 times improvement of success rate (agents reach goals and never collide) on < 500 agents, and still maintain more than 50% success rate for > 1000 agents when other methods completely fail. 🧵 (3/7)
1
0
1
For the first work, we look at distributed collision-avoidance in large-scale environments using local information. We introduce graph control barrier functions (GCBF) for distributed control and use GNN to learn the GCBF and the controller: https://t.co/ef08z4jfN6 🧵 (2/7)
1
0
1
Excited to present two works at #CoRL2023, both at Poster Session 6 on Thursday 4:15 pm! 🧵 (1/7)
1
0
3
An @MITAeroAstro team developed a machine-learning technique that can autonomously fly a plane through a difficult “stabilize-avoid” scenario while avoiding obstacles. Read more at MIT News https://t.co/9xouNpnWC0
2
12
41
Come to the poster sessions and oral presentations! Yue’s work will appear in Poster Session 1 while Songyuan’s will appear in Poster Session 2. Songyuan will provide a deeper introduction to his work on Friday at the “Oral Presentations 4” time slot. See you there! (2/2)
0
0
3
REALM has two strong works being presented at @l4dc_conf this year. @YueMengTHU 's “Hybrid Systems Neural Control with Region-of-Attraction Planner” and Songyuan Zhang’s “Compositional Neural Certificates for Networked Dynamical Systems” will be presented! To learn more... (1/2)
1
1
8
Please attend @l4dc_conf, if you're interested in hearing from Yue first-hand!
0
0
1
And further information can be found here:
1
0
1
You can read about the method on ArXiv here: https://t.co/NU9s0UzzxP
1
0
0
The method finds “[a] neural network Lyapunov function and a neural network controller to ensure the states within the region of attraction (RoA) can be stabilized.” Also, it runs in 1/4 of the time that other learning-based methods in the literature do!
1
0
0
A huge number robotics problems can be modeled as hybrid dynamical systems, and these are known to be difficult to define controllers for. In @YueMengTHU ‘s new work (accepted to L4DC), we have developed a method for designing controllers for this important class of systems!
1
0
7
enough soft robot simulator. (3/3) https://t.co/j8nYv6X0UI
Sound interesting? Then, play around with his implementation of this work which is located in REALM’s Architect repository:
0
0
0
Context: REALM has established its passion for differentiable simulation in several different results (see our work on Architect) and we were having a spirited discussion last week about how better sensor placement/design might be possible if we had a good (2/3)
1
0
0
An exciting question that came from a vigorous discussion at REALM last week: What differentiable simulators exist that can model soft robotic systems interacting withe the world? (1/3) #AcademicChatter #RoboticsChatter #DifferentiableSimulation
1
0
1