Robots now have to run the experiment, not just move the beaker: 30 physics tasks test acting to know, and the best agent succeeds 43.3% of the time

For the first time an embodied agent is placed in a setting where acting is for knowing: the goal is not to move an object to a specified place but to measure a hidden physical quantity, predict a response, or configure an apparatus to meet a specification. Earlier manipulation benchmarks state the state the scene should reach, and laboratory benchmarks check whether protocol steps were completed; scientific-discovery benchmarks do ask the agent to find a hidden law, but an experiment is a text command or a choice of input values and its result returns as text or an exact number, with no physical action in between. The 30 experiments span mechanics, electricity, optics, thermal physics and chemistry on 14 apparatus platforms, and the system parameters are resampled for every instance, so the answer cannot be known in advance yet can be scored exactly; every experiment is verified solvable by a reference solution that acts through the public interface and takes every reading from the instruments, without a language model and without access to the hidden parameters.

Scoring looks past the final number: the result must lie within an instrument-set tolerance, rest on readings recorded in the same episode, and leave the bench restored. Manipulation benchmarks score the final state and protocol benchmarks score completed steps, and neither asks what the agent's number rests on. At submission the environment seals the record and checks that every cited reading was produced by an instrument in this episode, that the measurement conditions held such as a tray at rest or a settled temperature, that the result agrees with a refit of the agent's own cited data, and that every part the objective does not use is back in its designated place; tolerances run from 0.2% to 5%, the tightest the instruments support given the reference error.

The interface comes in three nested levels that take over more of the acting, used to locate where failure lies. If failure came mainly from reasoning, handing the agent perception and motion planning should lift success substantially. T0 gives three camera images, robot state and seven-joint position targets; T1 adds promptable segmentation, ideal depth and collision-free Cartesian hand motion planned by cuRobo; T2 adds locate, grasp, carry, place, press, push and slide skills distilled from the strongest agents' successful runs. Success changes by only 1.7 to 10.0 points from T0 to T1, while T2 raises the three weaker agents by 10.0 to 26.7 points, yet with those skills they reach at most 14.3% in long-horizon multi-stage work.

Failures concentrate in the physical loop rather than in reasoning over readings already in hand. The stopping point differs between stronger and weaker agents: the strong ones mostly have sufficient readings, while the weak ones drop out before their readings suffice. The three strongest agents pass L1 to L4 at rates from 0.81 to 1.00, and their lowest stage pass rates are at knowing the result within tolerance (0.73 to 0.82) and restoring the bench (0.55 to 0.79); 60% of DeepSeek runs, 67% of Kimi runs and 90% of GLM runs stop while changing or observing the apparatus state. Attributed to the first decisive error, acting, observing and restoring together account for 75% to 100% of every agent's failed runs, and knowing for at most 25%.

The evaluated setting is a simulated teaching laboratory in Isaac Sim with a Franka Panda arm, a two-finger gripper and three cameras, covering 30 quantitative experiments on 14 apparatus platforms; each of the six models runs in the coding agent its developer ships, at the middle reasoning setting, with the same task card, budget and wall-clock limit. T0 and T1 are the evaluation levels, while T2 is a diagnostic level whose skills come from successful runs on the same experiments, so a T2 run partly reuses known solutions and is not part of the released protocol.

The authors state that T2 skills are distilled from the strongest agents' successful runs on the same experiments, so T2 runs partly reuse known solutions and serve only to measure how much of the failure lies in acting; T2 is not part of the released protocol. Evaluation is in simulation, where instruments carry quantization and noise models, and the findings are not yet validated on real laboratory benches; 3 of the 30 experiments are solved by no agent at T0, and Claude Opus 5.5 solves those three only at T1.