Robots Take On Open-World Tasks as Embodied AI Debate Turns to World Models

The robots were given an end-to-end chain: taking orders, analyzing customer needs, surveying the site, sorting ingredients and tools, grilling, turning food to control temperature, arranging finished items and delivering food. Five capabilities were repeatedly tested: active perception, multimodal understanding, streaming memory, dynamics modeling and continuous self-evolution. The service robot once brought tissues without being asked; when a customer asked for less spice, it understood the request and passed it to the grilling robot; when asked for water mid-order, it fetched the water and then resumed the order. The grilling robot failed its first flip, adjusted its grasp and succeeded on the second. LeEnjoy founder Guo Renjie showed the back kitchen and a small room behind a curtain to argue there was no space for teleoperation. He said the team had rented the site for tens of thousands of yuan after talking with the restaurant owner only on Sunday.

The live test followed a public dispute. In July, LeEnjoy launched the Etai model website and published its technical route. On Aug. 26, a 10-minute single-take demo of two robots cleaning a room circulated in embodied AI circles. On Sept. 3, OpenAI launched a GPT-6 Astra website; three days later, its chief scientist Jakub Pachocki published a long essay, “Alien Mind.” On Sept. 10, Guo published three questions to OpenAI, saying the technical ideas, wording and website design were highly similar, accusing OpenAI of “pixel-level copying” and “direct distillation,” and saying his legal team had started procedures. Public opinion split: some called it hard capability, others called it attention-seeking or a fake demo. LeEnjoy then put its robots on stage to respond.

According to Leiphone, Etai’s core mechanism is Energy-Driven, supported by neuron allocation. It integrates task completion, physical constraints, memory state and action feasibility into one energy function; lower total energy means a better state, and the robot chooses actions that reduce it. LeEnjoy contrasts this with VLA and WAM. VLA predicts action distributions and can suffer from information collapse, where language is ignored; a paper at ICML 2026, LangForce, named this problem, and Leiphone reported that on the RoboCasa benchmark a vision-only model performed almost as well as one receiving full language instructions. WAM predicts future world states and actions, adding physical prediction but still weak in semantics and costly: Nvidia’s DreamZero takes 25 days on eight H100 GPUs for one training run. LeEnjoy argues that prediction-based models fail outside their training distribution, while a scalar energy function can combine task goals and constraints without retraining for each new constraint.

On Sept. 17, Figure took Helix 2.5 into 30 unfamiliar homes in the San Francisco Bay Area to tidy living rooms, fold towels and make beds. According to QbitAI, Figure’s “zero-shot” claim means the robot did not retrain on the new home layouts or object instances; the tasks had been adapted earlier with data from other environments. In one experiment, two policy models differed mainly in whether they started from Index pretraining. The randomly initialized model achieved 9 percent full-task success, while the Index-pretrained model reached 56 percent; a task counted as successful only if fully completed, and human intervention for safety failed the episode. In a separate offline test using four nested Index data subsets spanning an eightfold range, Figure reported that validation loss for robot action prediction fell regularly as data increased, and that its prediction for the largest model was within 0.54 percent of the observed loss change across the range.

Figure has not published the closed-loop success rates corresponding to 1x, 2x, 4x and 8x Index pretraining, according to QbitAI, so the results do not show that real-world success rises at the same rate. Figure is building its own human-data supply chain. Through its Creator network, people record tasks in real homes and workplaces. By the end of August, Figure said it had received more than 16 million videos and paid creators $15 million; when Helix 2.5 launched, new human video was entering Index at about 35 minutes per second. The data go through quality filtering, anti-cheating checks, deduplication, rebalancing and text annotation. In early September, Figure announced a strategic partnership with British AI cloud provider Nscale involving an initial $3.5 billion compute commitment, potentially expanding beyond $6 billion, with plans to deploy up to 100,000 GPUs based on Nvidia’s Vera Rubin platform. Founder and CEO Brett Adcock said Helix’s next phase is increasingly constrained by data and compute. Figure has not disclosed the shooting-angle composition of the full Index dataset or a complete human-action supervision mechanism.

Other teams are testing human video pretraining. In August, Dyna Robotics released Dyna-2, using first-person human videos in four strictly nested scales: 1,000, 10,000, 100,000 and 1 million hours. After the same robot post-training, the four models reached 20, 28, 45 and 53 percent of task-reachable upper bounds across 14 real robot tasks, according to QbitAI; these normalized scores are not success rates and cannot be directly compared with Figure’s 56 percent. Dyna’s ablation found that predicting actions alone overfit badly, while adding future video prediction improved transfer across 39 robot tasks, and adding human video without action labels for video prediction produced stable monotonic gains in cross-embodiment performance as data scale increased. On Sept. 21, Light Origins published a technical blog on Light-O1, which recovers structured human actions from video, encodes them into a unified humanoid action representation, and pretrains autoregressively with vision and language to obtain a Human Action Prior. Its human-action pretraining scale grew from 3.75 billion to 120 billion tokens, corresponding to about 100,000 hours of human action at the largest scale, and was adapted to first-person human data, public robot data and its own humanoid robot data. CoMo, HuRo and Rhoda are also trying to extract transferable signals from internet video through representation learning, pseudo-actions or causal video models and inverse dynamics.

On the research side, Yann LeCun used his ECCV 2026 keynote, “World Models: Enabling the next AI revolution,” to argue that language and token training alone cannot reach human-level intelligence or cross the physical-world threshold. According to Leiphone, LeCun described four competing world-model routes: video generation from Sora and Genie betting on emergence; World Labs locking in 3D consistency; Nvidia Cosmos and Omniverse using explicit physics engines as robot training grounds; and JEPA, which predicts in abstract representations rather than generating futures. LeCun called pixel prediction a false problem because the next frame depends on information outside the current frame, such as whether someone enters or the light changes. He compared the roughly 10^14 bytes of public text available to large language models with the roughly 10^14 bytes a four-year-old receives through vision and touch, and said AI still lacks System 2-style reasoning. He advised researchers to stop focusing on generative models and LLMs and to turn to joint embedding, abstract representations and world models. His alternative uses an energy function to score candidate answers, with lower energy indicating a better answer.