Pretrained vision-language models carry rich common-sense priors, but turning them into controllable robot policies is hard. We introduce Steerable Policies: vision-language-action models trained on synthetic commands at multiple levels of abstraction — subtasks, motions, and grounded pixel coordinates — so that pretrained knowledge can be steered toward precise robotic behavior. Using both a learned embodied reasoner and prompted off-the-shelf VLMs, Steerable Policies improve generalization and long-horizon performance on real-world manipulation tasks.

How do we collect an internet-scale robotics dataset? The logistics of real-world data collection make scaling difficult, requiring expensive robot hardware, physical environments, and time-consuming scene resets. DexHub is a simulation-based robotics crowdsourcing platform that only requires a VR headset. Compared to alternatives, non-experts can collect data x2 faster with DexHub. Leveraging the flexibility of simulation, policies trained with DexHub are more robust than their real-world counterparts. Try it with Apple Vision Pro!

Evolution Gym is the first benchmark tool for studying soft robot co-optimization: jointly optimizing a robot’s body and brain. We built EvoGym to have 1) A fast soft-body simulator 2) OpenAI Gym inspired python interface and 3) over 30 environments in order to overcome the prior intractability and irreproducibility of soft robot co-design experiments. In addition, we proposed 3 SOTA co-design algorithms and benchmarked them extensively on all EvoGym environments to establish baseline performance in this field. We hope EvoGym will lead to the development of increasingly intelligent robots! Code and extensive docs are available!

Perceptual hashes map images with identical semantic content to the same n-bit hash value, and otherwise assign images to different hashes. Apple’s NeuralHash is one such system aiming to detect illegal content on users’ devices without compromising consumer privacy. We make the surprising discovery that NeuralHash is approximately linear, which inspires the development of novel black-box attacks that can (i) evade detection of ‘illegal’ images, (ii) generate near-collisions, and (iii) leak information about hashed images, all without access to model parameters. These vulnerabilities pose serious threats to NeuralHash’s security goals; to address them, we propose a simple fix using classical cryptographic standards.

We design and analyze novel, simple, and efficient algorithms for interactively learning non-binary concepts in the learning from random counter-examples (LRC) model. Here, learning takes place from random counter-examples that the learner receives in response to their proper equivalence queries, and the learning time is the number of counter-examples needed by the learner to identify the target concept. Such learning is particularly suited for online ranking, classification, clustering, etc., where machine learning models must be used before they are fully trained.