Aether AI (Causal Intelligence)

Towards Real-World Causal Intelligence
Founded by @huang_biwei

[Aether AI (Causal Intelligence) @AetherLab_AI]

Today's AI is remarkably good at recognizing patterns. But in the real world, patterns can only take us so far.

When a robot acts, every action changes its environment. To operate reliably, it needs to understand not just what to do next, but what its actions will cause and how to adapt when reality doesn't go as expected.

At Aether AI, we're building AI that reasons about cause and effect.

We started with robotics, where causality is tangible and every action has a clear consequence. But the ideas we're developing extend far beyond the physical world.

Check out our new brand film which tells the story!

[Aether AI (Causal Intelligence) @AetherLab_AI]

Scientific discovery often involves exploring a vast space of hypotheses. Can causal reasoning help narrow this space before experimental validation?

At Aether AI, we’re building the infrastructure for Causal Intelligence, enabling AI systems to reason about cause and effect, evaluate competing hypotheses, and identify promising directions for further investigation.

From intelligent robotics to scientific discovery, the goal is not to test every possibility, but to use causal reasoning to guide exploration and prioritize experiments that can resolve critical uncertainties.

[Aether AI (Causal Intelligence) @AetherLab_AI]

CausalWAM unifies world prediction and robotic action generation through a shared representation of physical dynamics, using feedback from recent interactions to continuously refine its predictions and actions.

[Biwei Huang @huang_biwei]

We introduce CausalWAM (Causal World Action Model), a unified framework for predicting future world states and generating robotic actions. Built on CausalWM, CausalWAM integrates an Action Expert with the world-modeling expert, using Causal Chain-of-Thought to represent intermediate physical variables and state transitions that support both prediction and action generation.

To enable training across different robots, CausalWAM uses a unified end-effector action representation grounded in camera observations. Trajectories collected at different control frequencies are aligned to a common temporal grid. During execution, a bounded memory maintains recent observations and interaction outcomes to support continuous replanning.

After data curation, we retained 15.8k trainable hours from 19.5k raw hours.

- On LIBERO-Plus, CausalWAM reached 93.0% average success after fine-tuning on the four standard LIBERO suites, with no additional fine-tuning on LIBERO-Plus.

- On RoboTwin 2.0 Clean2Random, it achieved 74.5% average success after fine-tuning only on clean demonstrations.

- On make_toast in RoboDojo, it reduced mean planner API cost per attempt from $11.04 to $0.03 (excluding local CausalWAM inference), while improving task success from 20% to 50%.

We also deployed the pretrained checkpoint directly on novel RoboLab tasks and scenes without additional fine-tuning to show the generalization ability.

Full blog in the reply below.

[Aether AI (Causal Intelligence) @AetherLab_AI]

Video2World found a large gap between code that runs and a world that works. Across 222 reconstruction instances, coding agents reached build rates between 87.6% and 93.1%, while every automated system completed fewer than 26% of the tasks.

[Biwei Huang @huang_biwei]

Coding agents have recently demonstrated emerging capabilities in solving complex software engineering problems. But how close are we to building causal world models with coding agents?

We introduce Video2World, a benchmark testing whether coding agents can reconstruct executable simulators from embodied videos. It covers 222 instances from 189 clips across 39 task families, evaluated across nine coding-agent systems.

The gap between runnable code and functional worlds is striking. Build rates reach 87.6%-93.1%, yet no automated system succeeds on more than 25.5% of tasks.

Accurate geometry alone is not enough. GPT-6 Astra achieves lower object-shape error than Claude Fable 5.1 (0.69 vs. 1.00 cm), but lower task success (10.7% vs. 25.5%). This is a cross-system comparison, not a controlled experiment.

In a separate small intervention study, we moved objects or targets by up to 10 cm in previously successful pick-and-place reconstructions. At 10 cm, replaying the original motion failed in every case, while the original behavior programs still succeeded in 83% of Astra worlds and 92% of Fable worlds.

Technical blog and resources in the reply below.

[Aether AI (Causal Intelligence) @AetherLab_AI]

Short-horizon manipulation is only one step toward robot autonomy.

With RobotAgentBench, we explore what happens when robots must discover tasks, reason about context, and recover from failures in open-ended scenarios. We also evaluate how agent harnesses and recursive self-improvement (RSI) can help close the gap.

[Aether AI (Causal Intelligence) @AetherLab_AI]

Short-horizon manipulation is only one step toward robot autonomy.

With RobotAgentBench, we explore what happens when robots must discover tasks, reason about context, and recover from failures in open-ended scenarios. We also evaluate how agent harnesses and recursive self-improvement (RSI) can help close the gap.

[Biwei Huang @huang_biwei]

Robots can now perform an increasingly wide range of manipulation tasks. But how far are we from robot assistants that can operate autonomously in the open world?

We introduce RobotAgentBench (RAB), a comprehensive benchmark for evaluating open-world robot autonomy in realistic household environments.

We evaluate 8 VLA models, 3 WAMs, and diverse robot agent harness designs across five dimensions: task discovery, context reasoning, manipulation, failure recovery, and robustness. Four open-ended household scenarios test how these capabilities work together to fulfill high-level user needs.

What we find:
- Manipulation ≠ autonomy. Standalone VLA/WAM policies struggle with task discovery and open-ended scenarios. Agent harnesses substantially improve performance, but reliable open-world autonomy remains challenging.
- Hybrid Code-VLA/WAM Policy performs best. Combining learned policies with code-based control achieves the strongest performance among evaluated configurations.
- RSI further improves robot agents. Refining reusable tools and accumulating experience lead to additional gains across iterations.
- Best couple under our default agent setting: Antigravity CLI (Gemini 4 Argon) + InternVLA-A1.5. This pairing performs best among the planner-executor combinations we evaluated, achieving an open-ended score of 10.4, compared with 8.4 for fully code-based Codex (GPT-6 Astra) and 1.1 for standalone π0.5.

[Aether AI (Causal Intelligence) @AetherLab_AI]

The cat has a theory about that button, but the bowl has a limit.

Each press adds food. Causal reasoning follows that chain to one decision. Pull the plug.

#CausalIntelligence #MachineLearning

[Aether AI (Causal Intelligence) @AetherLab_AI]

💡 Learn how we build real-world causal intelligence: https://t.co/zHF659jxqF

[Aether AI (Causal Intelligence) @AetherLab_AI]

What does causality mean when AI acts?

As our founder @huang_biwei puts it, “Causality is the ability to understand the rules that govern how actions change the world, and to reason about their consequences before acting.”

The result of one action shapes the next decision.

#CausalIntelligence

[Aether AI (Causal Intelligence) @AetherLab_AI]

The path to Aether AI began long before the company itself.

Causal learning taught us to look beyond mere patterns and unlock the mechanisms that drive outcomes. Foundation models made it possible to do this at scale.

We’re building the infrastructure for Causal Intelligence, with a horizon that includes physical AI, computer use, forecasting, and scientific discovery.

[Aether AI (Causal Intelligence) @AetherLab_AI]

As our founder @huang_biwei shares, CRIS-0 brings our causal world model and causal agent together on a physical robot for the first time. Throughout the demo, one reasoning loop handles changing tasks and live disruptions, from a metal can on the plate to a hand in the door's path.

Physical AI is our first proving ground, and the same architecture lays the foundation for generalizing to very different domains.
https://t.co/RWhEYEaoNV

[Biwei Huang @huang_biwei]

I have spent more than a decade studying causality. For me, causality is about understanding the rules that govern how actions change the world, and reasoning about their consequences before acting.

Scaling has made today's models remarkably capable at spotting patterns and predicting what comes next. A system that acts in the physical world needs more. It has to ask what will happen if it acts, check the outcome, and adapt when reality doesn't match the plan. In the physical world, trial and error can be costly or unsafe, and an early mistake can carry through the rest of a task.

When we started Aether AI, robotics was the obvious place to begin.

CRIS-0 is our first integrated system built on these ideas. It tracks the conditions that matter for a task, predicts how each action will change them, and replans when something unexpected happens. In our demo, that means taking a metal can off a plate before heating it, and telling an empty can from an unfinished drink while tidying a room.

Physical AI is just our first proving ground. We're building towards causal intelligence that also works for forecasting, scientific discovery, and more.

The technical write-up is linked in the comments.

[Aether AI (Causal Intelligence) @AetherLab_AI]

@svpino Thanks! Excited to keep building!

[Aether AI (Causal Intelligence) @AetherLab_AI]

This is CRIS-0, Aether AI's first causal robotic intelligence system, operating in a real home. It reasons about how its actions will change the world, checks what actually happened, and adapts when reality doesn't go as expected.

[Aether AI (Causal Intelligence) @AetherLab_AI]

@GithubProjects Thank you for the support!

[Aether AI (Causal Intelligence) @AetherLab_AI]

@_jaydeepkarale Exactly. That's a key part of what we're building.

[Aether AI (Causal Intelligence) @AetherLab_AI]

@Whats_AI Thank you! We’re really excited about this one

[Aether AI (Causal Intelligence) @AetherLab_AI]

@VoidAsuka Thanks! Excited to share CRIS-0 with everyone

[Aether AI (Causal Intelligence) @AetherLab_AI]

A failed step should not derail an entire task.

Built on Aether AI's causal world model and causal agent, CRIS-0 is designed to make long-horizon tasks more reliable. It breaks complex tasks into stages with observable completion conditions, verifies outcomes before moving forward, and adapts when reality diverges from expectation.

When something goes wrong, CRIS-0 can retry, roll back to an earlier stage, or revise its plan, rather than blindly continuing down the wrong path.

Read the full piece on TechCrunch: https://t.co/ddRxgeH2tT

[Aether AI (Causal Intelligence) @AetherLab_AI]

@lyndsyfonsecaX That’s the point! CRIS-0 takes a broad request and works out the next steps based on each object’s state.

[Aether AI (Causal Intelligence) @AetherLab_AI]

@MayaLeClarke If the current step can be retried, it tries it again. If that’s not enough, it replans based on the updated task state.

[Aether AI (Causal Intelligence) @AetherLab_AI]

Under the hood, CRIS-0 predicts how actions can change the state of the world, plans around those predictions, and checks what actually happens to inform its next move.

Predict. Act. Verify. Adapt.

Full write-up: https://t.co/9O5Jf758L3

[Aether AI (Causal Intelligence) @AetherLab_AI]

We asked CRIS-0 to tidy the living room.

Some cans were empty. Others still had drinks in them. CRIS-0 kept track of their states, threw away the empties, and set the unfinished drinks aside.