@AnsonISL

DPhil student @a2i_oxford | Machine Learning, Causality, Robotics

 Oxford, England
Joined March 2020
  • Tweets18
  • Following178
  • Followers119
7Media
We’ve been thinking a lot about what it means for a model to understand the world, and prediction alone doesn’t seem to be the full answer. This is our attempt to write down our thoughts on AI and scientific discovery, and how mechanistic structures might be the answer.
Ingmar Posner@IngmarPosnerJul 16
Introducing: Mechanistic World Models. Scientific discovery is not just prediction. It requires uncovering the mechanisms generating what we observe. A capability current ML lacks. With @AnsonISL and @bschoelkopf. @a2i_oxford, @oxengsci and @MPI_IS. arxiv.org/abs/2607.12474
1
10
749
Introducing: Mechanistic World Models. Scientific discovery is not just prediction. It requires uncovering the mechanisms generating what we observe. A capability current ML lacks. With @AnsonISL and @bschoelkopf. @a2i_oxford, @oxengsci and @MPI_IS. arxiv.org/abs/2607.12474
9
52
5
311
22,585
I am excited to introduce GLAM (Grounded Latent-Action World Models), our latest effort to enable robot learning on heterogeneous data using world models. Website + videos: viccccciv.github.io/glam/ Paper: arxiv.org/abs/2606.21672 1 / N 🧵👇 #robotics #machinelearning #worldmodel
4
26
2
111
11,708
How can a bimanual robot grasp objects that are initially ungraspable due to environmental constraints? COMBO-Grasp tackles this challenge by leveraging a constraint policy to stabilise the object while a grasp policy executes non-prehensile manipulation and grasping.
3
18
3
168
16,481
🚀 Funded PhD Studentship Alert! 🚀 Join @a2i_oxford in our mission to advance World Models for robotics and beyond. Two opportunities starting Oct 2025. Deadlines in January 2025. Details in 🧵… @oxfordrobots 🤖 #Robotics #ArtificialInteligence #MachineLearning
5
20
128
24,693
Very excited to share our new work - SPARTAN: A Sparse Transformer Learning Local Causation. We develop a Transformer world model that learns local causal dependencies between entities, leading to improved adaptation efficiency and robustness with accurate prediction. 🧵
2
2
2
15
3,399
more replies
As a result, SPARTAN achieves significantly improved efficiency when adapting to interventions as well as robustness against distractors and changes to irrelevant objects, without compromising on prediction accuracy.
1
3
151
Check out the paper for more details! w/ @IngmarPosner @bschoelkopf Work done at @a2i_oxford arxiv.org/abs/2411.06890

SPARTAN: A Sparse Transformer World Model Attending to What Matters

Capturing the interactions between entities in a structured way plays a central role in world models that flexibly adapt to changes in the environment. Recent works motivate the benefits of models...

arxiv.org
3
125
🚀 We’re hiring a Postdoc in Multimodal World Modelling for robot skill acquisition! 🌟 Do you have a passion for deploying AI on real-world robots? Then this one may be for you... my.corehr.com/pls/uoxrecruit… #robotlearning #Robotics #GenerativeAI #MachineLearning #AI @oxfordrobots
2
1
10
3,760
My group in @oxengsci, @oxfordrobots is looking for talented research students passionate about robot learning. Interested in doing a PhD researching efficient and versatile world models for robotics and beyond? This one may be for you…
Oxford Applied AI Lab@a2i_oxford12 Feb 2024
If you are excited about world-models in robotics, check out this EPSRC iCASE PhD studentship (fully funded for UK Home students) in @a2i_oxford in collaboration with Siemens: Foundation Models for Industrial Control Applications. *Deadline: 1 March 2024* ox.ac.uk/admissions/graduate…

DPhil in Engineering Science

A research-based course offering advanced study in areas like biomedical, energy, materials, civil and information engineering, with interdisciplinary projects and strong industry collaboration.

ox.ac.uk
2
10
19
9,190
Excited to present AMP-LS, our recent work on gradient-based motion planning, accepted to @ieee_ras_icra 2023. AMP-LS can plan a collision-free trajectory efficiently in novel complex scenes and is trained on only easily generated kinematically valid joint states. 🧵👇
1
2
5
20
5,227
Our 2012 paper ‘On causal and anticausal learning’ just received a Test of Time Honorable Mention at @icmlconf #ICML2022: icml.cc/2012/papers/625.pdf. I am really grateful, and would like to use this occasion for some thoughts on causality and machine learning:
7
147
7
982
Super excited to share our new work - a personal first - Variational Causal Dynamics (VCD). We trained a latent state space model with a causal and modular structure for efficient adaptation. Work at @a2i_oxford with @bschoelkopf, @IngmarPosner Paper: arxiv.org/abs/2206.11131 1/3
1
25
2
111
VCD learns a causal transition model by performing causal discovery in the latent space. The key is to learn the representation and the transition model **jointly** under sparsity regularisation. In doing so, we get a nice bi-product: disentangled latent representation. 2/3
1
1
8
The end result? With a small amount of observations from an unseen environment, VCD can identify sparse changes in the scene and adapt modularly. This results in significantly improved sample efficiency during adaptation. 3/3
1
8
Load more