# ANTS

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Alfredo Wijaya Hardjoprawiro](https://cerebralvalley.ai/u/alfredowh), [Daniel Totolici](https://cerebralvalley.ai/u/siliconcowboy)
- **GitHub:** https://github.com/alfredowh/ANTS
- **Demo video:** https://www.youtube.com/watch?v=F6niehGrYgg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5136

ANTS: Adversarial Agentic Task Synthesis

Show a robot a task once. Then let Meta's models try to break it, and turn every failure into a line of code you can read.
ANTS turns one demonstration into a robot skill, then hardens it by attacking it. Meta's reasoning models and their understanding of the physical world drive all three parts of the system: the synthesizer, the adversary and the orchestrator. No model is trained or fine-tuned. All improvement happens through context.
The problem

Robot policies are brittle. A policy that scores well on a benchmark can fail after a small change in camera angle, object pose or friction. Each new task also needs new data and fine-tuning. A factory can't certify a skill whose failure modes are unknown.
Synthesizer: from one demonstration to a program
Muse Wonder reasons over the demonstration as a physical task: what is grasped, where it moves and in what order. It writes the skill as a short, readable robot program, not network weights.
Adversary: finding what breaks

Muse Spark agents read the program and look for the physical assumptions it hard-codes, such as a fixed object position, a straight approach path or no re-check after a grasp. They propose attacks in simulation (LIBERO, Franka Panda), where SAM 3 provides the robot's perception. A Thompson-sampling planner spends the test budget on the categories most likely to fail, and a hard severity bound in code keeps every attack realistic. Muse Spark then diagnoses each failure, links the physical cause to the program and states a constraint the generator can follow.

Orchestrator: closing the loop
Each failure goes back as feedback, and Muse Code or Muse Wonder rewrites the program under the loop's written workflow. The safety decisions stay in deterministic code, not with the model. A regression gate keeps a rewrite only if it is no worse on every case seen so far and better on at least one, so robustness never regresses. If the generator stalls on the same failure, ANTS asks a human for one targeted demonstration.

A worked example
Take a bowl-on-plate task. The first program copies what the demo showed and passes its nominal replay. A bowl that has been moved makes it miss, so the program learns to read poses from perception at run time. Clutter on the path causes a collision, so it learns to approach with clearance. A shifted camera causes another miss, so it learns to re-localise and retry. Each added line traces to one diagnosed failure.

Why it matters
Every fix comes with a replayable audit trail. Black-box policies can't offer that. ANTS shifts the question from "can the robot do it?" to "can we prove it keeps working?"

Evaluation
We tune only on an acceptance set and report on a separate held-out set. Every number comes from a logged run.

Limits
One task and one robot, simulation only. Skills are limited by the primitive set, and only the perturbations we define are searched.

## More from Meta Global AI Developer Hackathon

- [Dueboard](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5133)
- [Intentional Scrolling](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5134)
- [StartAppLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5135)
- [theoretical-incoherent-universals](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5137)
- [Big Bertha](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5138)
- [Hyperion Industries](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5139)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5136. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt