Robotics has reached the same point language models sat at during the GPT-2 era, according to Perry Dong: fluent, capable pretrained systems that cannot be trusted to act alone. Dong, writing on his personal blog alongside Chelsea Finn, points to labs including Physical Intelligence, Generalist, and Google DeepMind, all of which have shown pretrained robot models handling complex tasks. Capability alone, he argues, does not make a robot safe to deploy. A dishwashing robot that succeeds 95 percent of the time will still break something roughly once a week in a house full of glass, pets, and small children.
Language models solved this exact problem, Dong writes, through a standardized process: start from a pretrained model, define rewards, run a known family of reinforcement learning algorithms, and watch for specific failure modes such as reward hacking. Any team can follow that documented playbook to fine-tune a large language model. Robotics has no equivalent. Getting reinforcement learning to work on a robot today is “a craft,” not a recipe, limited to a handful of specialists who tune systems by feel.
The gap is structural. Systems that already work at scale, such as the reinforcement learning behind AlphaGo or the method used to fine-tune chatbots, depend on actions that are cheap to generate and easy to verify. A Go move or a text response can be scored the moment it appears. A robot folding laundry cannot try a thousand actions per step, and a single task can chain together thousands of low-value control commands before any reward shows up. Samples are expensive and the payoff arrives only at the end. That combination pushes robotics toward algorithms that learn a scoring model and lean on it to work out, over a long stretch of steps, which of the earlier moves actually mattered. It is an approach with far less track record at scale than the on-policy methods used for language models.
To test whether stable training is even possible, Dong built EXPO-FT, a reinforcement learning method that generates several candidate actions from a pretrained robot model, edits each one slightly with a smaller learned policy, and keeps whichever edit scores highest on a value function. The edits stay deliberately small, so the system cannot push the robot into an unsafe action, and whatever improvement the edits find gets folded back into the base model over time. Dong reports 30 successes out of 30 attempts across six manipulation tasks, among them routing string lights through hooks and setting a flower inside a bottle’s neck, needing roughly 19 minutes of real interaction time per task on average. Those are the authors’ own reported results from real robot trials, not an independently verified benchmark.
An algorithm alone does not make a recipe, Dong is careful to note. Robotics post-training still needs standard answers to questions LLM fine-tuning already settled: how to define success automatically instead of having a human watch every attempt, how to reset a robot’s environment between tries without a person doing it by hand, and how much human intervention a training run needs as it scales from one robot to a fleet.
For robotics labs still leaning on manual reward design and constant supervision, EXPO-FT reads less like a finished tool and more like proof that stable large-model reinforcement learning is possible at all. The bigger unlock, standardized reset and reward protocols, is still unbuilt, and whichever lab defines that layer first sets the default the rest of the industry copies.
Perry Dong published this analysis on his personal blog in September 2026.
