Two A100 GPUs and a Video World Model: How Google and NYU Want to Cut Robot Training Costs

Every real-robot training run is, at heart, a small act of destruction. A manipulation trial takes hours of wall-clock time, wears out hardware, and the results barely transfer between labs because every rig is calibrated differently. That cost ceiling is the reason embodied AI, the field of AI that controls physical robots rather than chats through a screen, has not had its ImageNet moment. Sherry Yang (homepage), a researcher at NYU and Google DeepMind, told attendees at the Robotics: Science and Systems conference (RSS 2026) that the workaround is to stop rehearsing on metal and start rehearsing in video.

Yang's talk, titled "Evaluating and Improving Physical Agents in a World Model," anchors a research thread that has produced three independent artifacts in five months. The first is WorldGym, an action-conditioned video world model published in June 2025 (arXiv:2506.00613). WorldGym is an autoregressive video generator: feed it a single start frame and a sequence of robot actions, and it dreams the rest of the episode. It is small enough to train on two NVIDIA A100 GPUs, per the WorldGym project page, and the team reports that its in-world success rate correlates with real-world success rate, making it a proxy environment for policy evaluation. The "two A100" cost figure is a talk-stage claim from the RSS 2026 session; the paper's compute footprint is consistent with that scale but is not independently re-benchmarked.

The second artifact is World-Gymnast, released in February 2026. It fine-tunes a vision-language-action (VLA) policy, a model that maps camera input directly to robot joint commands and serves as the embodied-AI counterpart of an LLM, with reinforcement learning. The rollout happens inside an action-conditioned video world model, not in simulation software and not on a real robot. A vision-language model scores each imagined trajectory. On the Bridge benchmark, a standard low-cost robot-arm manipulation testbed, World-Gymnast outperforms supervised fine-tuning by as much as 18x and software-simulator RL by as much as 2x, where "as much as" is the best-case figure across tasks, not an average. Held out on OpenVLA, World-Gymnast reports SFT at 58 ± 4%, World-Gymnast at 74 ± 3%, with distractors, language grounding, and scaling lifting the headline number to 81 ± 4%. Test-time RL, starting from a single novel frame, lifted a close-the-drawer task from 62 ± 6% to 100 ± 0%.

The third artifact is a position paper, "World Models as an Intermediary between Agents and the Real World" (arXiv:2602.00785, January 31, 2026), which frames the world model as a dynamics-plus-reward-plus-task-distribution stand-in for the physical world across robotics, ML engineering, computer use, and science. Together the three pieces make the same argument: in embodied AI, the constraint is no longer who has the best VLA. It is who can afford to rehearse it.

When Yang's group dropped RT-1, an earlier Google robot policy, into their world model, RT-1's control actions went so far out of distribution that the world model visibly broke. That is a readout of which policies the rehearsal surface can and cannot evaluate. Out-of-distribution image-editing probes pushed further: a carrot drawn inside the computer monitor on the desk, a rubber duck placed next to the gripper, a toy car blocking the workspace. The carrot-on-monitor probe is the one that landed at RSS; the model kept "treating" the carrot as if it were part of the real scene, exposing a binding failure that a real-robot evaluator would have caught only by accident.

There is also a structural finding about what VLA policies actually learn. The team reports a color-vs-shape gap: when the same object appears in an unexpected color, success drops sharply; when it appears in an unexpected shape, the drop is smaller. That asymmetry is not obvious from a benchmark score, and it is the kind of diagnostic the field will need to take embodied models past controlled demos.

A video world model is a rehearsal surface, not a replacement for the world. The "two A100 GPUs" line sets a democratization floor, not a ceiling; the same paper notes that the RL pass and the VLM judge add their own compute costs. The 18x and 2x comparisons are best-case across tasks, and the RSS audience saw them on a talk slide. The position paper is, by its own admission, a frame, not a result. The Leiphone RSS 2026 on-site writeup by Zhang Lu is original reporting from the room, but the underlying claims still trace to Yang's group, not independent re-runs.

What changes if the workaround holds is the question of who can train an embodied policy at all. A rehearsal surface that fits on two A100s and runs against a video model rather than a hardware rig moves the bottleneck from robot hardware to machine-learning engineering, which is a bottleneck a lot more teams can clear. The RSS session ended without a paper-level commitment, but Yang said a 70B-scale world-model run is on the team's near-term roadmap. That is the next data point worth watching.