Single Agent vs Grasp Delegation

The same seven RoboLab pick-and-place tasks on a Franka DROID arm, three repetitions each, seed 0, DeepSeek policy. Single agent: one Native agent with direct set_gripper control (EXP10A baseline, commit 85a14ac1be16). Delegated: the parent has no set_gripper and hands each object's pick-and-place to a fresh child agent through grasp.engage (config <task>-pickplace-v1, commit 79e1421). Every verdict is the environment's own.

Success by task

Scores are successes over valid runs. Undecided runs count as valid non-successes on both sides. The one infrastructure-invalid delegated run is excluded.

Side by side

Every repetition

Choose a repetition to load both runs into the player above. Single-agent attempts are in time order. Repetition n on one side has no special link to repetition n on the other; they are simply the n-th attempt of each.

How to read it

• The two batches ran at different Harness commits with different tool surfaces. This is a side-by-side of two diagnostic batches, not a controlled A/B test or a benchmark success rate.
• Undecided: no official verdict. For delegated runs, the parent called finish_episode before the environment decided. The single-agent archive calls these "No verdict".
• Invalid: delegated FoodPacking1BoxesTask repetition 3. A child's model request was rejected as too large (HTTP 413), and the resumed parent then acted twice on one observation, which the trace verifier forbids.
• The single-agent archive's scripted tool probes are left out. They exercise tools and have no verdict.
• Each video shows the observation frames recorded at decision boundaries. Motion between observations was not captured continuously, so durations are not comparable as wall time.