Uploaded November 2026 | Updated November 2026, 3 hours ago
Prerecorded presentation at the 6th Applied Active Inference Symposium (November 12–13, 2026), hosted by the Active Inference Institute. Online and free.
Matthew Brown: “Unlocking Robotic Dexterity”
Most real-world work is messy, variable, and full of contact: cables that flex, parts that don’t quite fit, valves that stick. Traditional robots excel only where conditions are controlled and repeatable, and today’s leading AI approaches (vision-language-action models, diffusion policies, and embodied foundation models) struggle to close the gap. They depend on massive datasets that don’t exist for physical interaction, require heavy compute at inference time, and break down when reality drifts from their training distribution.
This talk argues that dexterity is not primarily a data problem but an adaptation problem. Nature doesn’t solve manipulation by training once offline and hoping it generalizes. It learns continuously at the point of contact, with perception, action, and learning operating as a single loop. We contrast the dominant paradigm of universal function approximation, which is retrospective, data-dependent, and time-agnostic, with enactive synchronization: a prospective, time-aware process in which an agent continuously adjusts to maintain predictive coherence with its environment.
Tracing a line from W. Ross Ashby’s cybernetics and his Homeostat through the Free Energy Principle and Active Inference, we show how ultrastable networks can form a minimum-prediction-error synchronization manifold between a robot and the world it touches. Sensors and motors become the interface through which goals are expressed as preferred states, and behavior emerges from the network’s drive to stay in equilibrium.
We then show what this looks like in practice: compact models that run on a single CPU at the edge, update in under a millisecond, and learn tasks in minutes rather than weeks, illustrated through new demonstrations of adaptive dexterous manipulation on real robotic hardware.
Symposium program: https://2026symposium.activeinference.institute/
Abstract book (Zenodo): zenodo.org/records/23986505
Livestreams: Day 1 youtube.com/watch?v=-v4IaFY9MZI · Day 2 youtube.com/watch?v=ccze9ay8D0o
All Symposium videos (playlist): youtube.com/playlist?list=PLO2noumu0LRk
Previous Symposia: https://symposium.activeinference.institute/
Active Inference Institute information:
Website: https://activeinference.institute/
X: https://x.com/InferenceActive
Discord: https://discord.activeinference.institute/
Donate: https://donate.activeinference.institute/
YouTube: youtube.com/c/ActiveInference
Active Inference Livestreams: https://video.activeinference.institute/
Prerecorded presentation at the 6th Applied Active Inference Symposium (November 12–13, 2026), hosted by the Active Inference Institute. Online and free.
Matthew Brown: “Unlocking Robotic Dexterity”
Most real-world work is messy, variable, and full of contact: cables that flex, parts that don’t quite fit, valves that stick. Traditional robots excel only where conditions are controlled and repeatable, and today’s leading AI approaches (vision-language-action models, diffusion policies, and embodied foundation models) struggle to close the gap. They depend on massive datasets that don’t exist for physical interaction, require heavy compute at inference time, and break down when reality drifts from their training distribution.
This talk argues that dexterity is not primarily a data problem but an adaptation problem. Nature doesn’t solve manipulation by training once offline and hoping it generalizes. It learns continuously at the point of contact, with perception, action, and learning operating as a single loop. We contrast the dominant paradigm of universal function approximation, which is retrospective, data-dependent, and time-agnostic, with enactive synchronization: a prospective, time-aware process in which an agent continuously adjusts to maintain predictive coherence with its environment.
Tracing a line from W. Ross Ashby’s cybernetics and his Homeostat through the Free Energy Principle and Active Inference, we show how ultrastable networks can form a minimum-prediction-error synchronization manifold between a robot and the world it touches. Sensors and motors become the interface through which goals are expressed as preferred states, and behavior emerges from the network’s drive to stay in equilibrium.
We then show what this looks like in practice: compact models that run on a single CPU at the edge, update in under a millisecond, and learn tasks in minutes rather than weeks, illustrated through new demonstrations of adaptive dexterous manipulation on real robotic hardware.
Symposium program: https://2026symposium.activeinference.institute/
Abstract book (Zenodo): zenodo.org/records/23986505
Livestreams: Day 1 youtube.com/watch?v=-v4IaFY9MZI · Day 2 youtube.com/watch?v=ccze9ay8D0o
All Symposium videos (playlist): youtube.com/playlist?list=PLO2noumu0LRk
Previous Symposia: https://symposium.activeinference.institute/
Active Inference Institute information:
Website: https://activeinference.institute/
X: https://x.com/InferenceActive
Discord: https://discord.activeinference.institute/
Donate: https://donate.activeinference.institute/
YouTube: youtube.com/c/ActiveInference
Active Inference Livestreams: https://video.activeinference.institute/




![ActInf GuestStream 115.1 ~ Energy-Based Transformers and the Future of Scaling
Energy-Based Transformers are Scalable Learners and Thinkers
Alexi Gladstone, Ganesh Nanduru, Md Mofijul Islam, Peixuan Han, Hyeonjeong Ha, Aman Chadha, Yilun Du, Heng Ji, Jundong Li, Tariq Iqbal
https://arxiv.org/abs/2507.02092
Inference-time computation techniques, analogous to human System 2 Thinking, have recently become popular for improving model performances. However, most existing approaches suffer from several limitations: they are modality-specific (e.g., working only in text), problem-specific (e.g., verifiable domains like math and coding), or require additional supervision/training on top of unsupervised pretraining (e.g., verifiers or verifiable rewards). In this paper, we ask the question Is it possible to generalize these System 2 Thinking approaches, and develop models that learn to think solely from unsupervised learning? Interestingly, we find the answer is yes, by learning to explicitly verify the compatibility between inputs and candidate-predictions, and then re-framing prediction problems as optimization with respect to this verifier. Specifically, we train Energy-Based Transformers (EBTs) a new class of Energy-Based Models (EBMs) to assign an energy value to every input and candidate-prediction pair, enabling predictions through gradient descent-based energy minimization until convergence. Across both discrete (text) and continuous (visual) modalities, we find EBTs scale faster than the dominant Transformer++ approach during training, achieving an up to 35% higher scaling rate with respect to data, batch size, parameters, FLOPs, and depth. During inference, EBTs improve performance with System 2 Thinking by 29% more than the Transformer++ on language tasks, and EBTs outperform Diffusion Transformers on image denoising while using fewer forward passes. Further, we find that EBTs achieve better results than existing models on most downstream tasks given the same or worse pretraining performance, suggesting that EBTs generalize better than existing approaches. Consequently, EBTs are a promising new paradigm for scaling both the learning and thinking capabilities of models.
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2507.02092 [cs.LG]
(or arXiv:2507.02092v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2507.02092
Active Inference Institute information:
Website: https://www.activeinference.institute/
Activities: https://activities.activeinference.institute/
Discord: https://discord.activeinference.institute/
Donate: http://donate.activeinference.institute/
YouTube: https://www.youtube.com/c/ActiveInference/
X: https://x.com/InferenceActive
Active Inference Livestreams: https://video.activeinference.institute/ ActInf GuestStream 115.1 ~ Energy-Based Transformers and the Future of Scaling](https://i.ytimg.com/vi/85ob-jrG7xQ/mqdefault.jpg)





