Aether AI (Causal Intelligence)

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.

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.

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.

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.