We fine-tuned a separate π0.5 policy for each of the five tasks in the MetaFine Open Competition, with a task-specific training mixture and number of actions executed per prediction. Training started from the official demonstrations and expert procedures. Local evaluation exposed recurring grasping, pushing, and insertion failures. We addressed them by revising the demonstrated motions and by adding expert continuations from states the policy reached, while keeping nominal demonstrations in training. Checkpoints were compared on every instruction, so that a gain on one instruction could not hide a loss on another. Our submission, MAI, ranked first overall with a score of 96.66.
Figures & tables
Task
Observed behavior
Training approach
Evaluation checks
T1
Part-dependent failures; slips as lifting begins
Part-specific grasps; gradual lift during closure
Cap and body scored separately
T2
All nominal cases succeed; two perturbation failures
Balanced sources and motion directions
All directions; visual perturbations
T3
Missed contact and partial pushes at difficult angles
Complete pushes; recovery across approach angles
Both switch colors; reserve scene set
T4
Empty grasps near close neighbors
Neighbor clearance and empty-grasp recovery
All colors; standard and crowded scenes
T5
Letter o released offset or unseated
Continuous plans with consistent grasps; full seating
All four letters; two scene sets
Table 1: Observed behavior, training approach, and evaluation checks for each task.
T1
T2
T3
T4
T5
Overall
P
86.25
98.89
98.54
99.31
94.37
95.47
U
90.00
100.00
100.00
100.00
96.25
97.25
B
90.00
100.00
100.00
100.00
96.25
97.25
S
88.75
99.63
99.51
99.77
95.62
96.66
Table 2: Official results (%). Overall averages the five tasks.
Task
Evaluation scope
Reference
Selected
T1
80 scenes; n=3,200
88.95
93.39
T2
20 scenes; n=480
—
100.00
T3
120 scenes, four inference seeds; n=7,680
99.22
99.47
T4
80 scenes; n=1,280
99.81
99.97
T5
80 development scenes; n=880
95.07
97.95
Separate 80 scenes; n=880
93.26
97.05
Table 3: Local composite scores S (%): Selected denotes the final submitted policy; Reference denotes a preceding task-specific policy evaluated on the same scene set. T2 was trained from the base model; “—” means no reference comparison is reported. n counts episodes per model. For T2, S follows the public competition configuration, which evaluates P on the left instruction; the sweep also covers right and forward.