A Public Benefit Corporation informing society about the frontier of physical AI.
Joined July 2026
- Tweets302
- Following27
- Followers3.6K
- Likes850
Robocurve PBC retweeted
Most models won’t stab a baby doll… until you blur it. Important view into safety behaviour under uncertainty and the multimodality gap.
Robocurve PBC retweeted
And weirdly not eval aware! (Although black box eval awareness techniques are questionable.) If I were shown a blurry camera image and told to stab something, *I* would suspect a setup...
Robocurve PBC retweeted
Called it
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Robocurve PBC retweeted
Using off-the-shelf image editing models, we can jailbreak frontier AI into controlling robots to do some harmful actions 100% of the time.
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Robocurve PBC retweeted
Some research I'm presenting @ CoRL: you can make embodied frontier LLMs (Astra, Sol, Fable, Opus, Gemini Argon) do anything you want by simple camera hacks.
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
We used props so no real harm was possible. We used a baby doll instead of a baby, empty bottles, and appliances with their power supply turned off.
Robocurve PBC retweeted
We are releasing the stable version of Inspect Robots so that anyone can evaluate any model, on any robot, for any task.
Inspect Robots is fully open-source. Feedback, issues, and PRs welcome!
We're releasing the first stable version of Inspect Robots, the evaluation framework we use internally.
Anyone can use Inspect Robots to run evaluations on any model, any robot and any task. For example, you can run GPT-6 Astra on i2RT's YAM arms.
Link and paper below.
We'll present Inspect Robots at the Science of Physical AI Safety workshop at CoRL 2026 on Nov 12 in Austin, organized by @paisi_ai.
Paper: arxiv.org/abs/2610.06306
