RoboFoundry Lifts GPT-5.5 on EmbodiedBench From 56.9% to 72.7% Without Fine-Tuning

The same GPT-5.5 produces two different results on EmbodiedBench. Connected directly, it successfully completes 56.9% of tasks on average. With RoboFoundry running around it, that figure rises to 72.7%. The authors did not touch the model's parameters. The gain is 15.8 percentage points, or roughly 28% relative to the baseline. Jiqizhixin reports.

What the numbers show

EmbodiedBench evaluates multimodal models as agents for embodied tasks, where they must perceive an environment and plan actions. In tests like this, the result has two components: the model itself and the scaffolding around it. Usually, researchers compare different models inside the same scaffolding. Here the setup is reversed. The model is fixed, and the system around it changes.

Key figures from the available part of the material:

• model: GPT-5.5, weights unchanged;
• benchmark: EmbodiedBench, average success rate;
• direct use of the model: 56.9%;
• with RoboFoundry: 72.7%;
• absolute gain: 15.8 percentage points;
• relative gain: about 28%.

The original headline says robots gain the ability to evolve on their own. This implies that the system improves as it runs, rather than being tuned manually once.

Where the difference hides

The authors explain the gap simply: it arises outside the model. It comes down to what information reaches the model in the first place and what experience the system keeps for itself. These are two levers, and both belong to context engineering, not neural network training.

For an embodied agent, this framing makes sense. The model sees the world only through what it is given: a scene description, an action history, hints about state. If the prompt contains extraneous material, the model gets distracted. If something needed is missing, it fails for no good reason. And if the system does not remember failures and successful tricks, every new attempt starts from zero. The available excerpt does not disclose RoboFoundry's concrete implementation, so the details of how it selects and stores experience fall outside this analysis.

The takeaway for the industry is practical. A 28% gain without a single training step means that part of what frontier models can do in robotics is bottlenecked by the scaffolding. Improving it is cheaper than fine-tuning a model, and the result is not tied to one version of the weights.

What is still missing

The available text lacks several things that make the 72.7% figure hard to evaluate fully:

• a breakdown by benchmark subtask: the average may hide failures or record jumps in individual categories;
• the number of evolution iterations and the cost of requests, since a system that accumulates experience spends tokens and time;
• results for other models: it is unclear whether the approach works as well on weaker or open LLMs;
• tests on physical robots, rather than only in simulation.

For now, this is a result on a single benchmark, and it should be read that way.

What teams building agents should check

The practical lesson applies to anyone building agent systems, not just robots. Before switching to a more expensive model, it is worth measuring how much is lost at the system level. Three questions can help find those losses quickly:

• what exactly the agent sees as input, and whether it can be shortened without losing meaning;
• whether the outcomes of past attempts are saved and used in later ones;
• how progress is measured: on a fixed model, with a repeatable set of tasks.

If RoboFoundry confirms the result on other models and in a real environment, competition in embodied AI will partly shift from model size to the quality of the system around the model.