Retain play_stacking_toy and insert_tubes, with the toy ranked higher for difficulty and diversity. Both completed successfully in the clarified-instruction RLE test: 9,297 and 5,117 native frames respectively, with valid verdicts, complete provider usage and independently verified terminal geometry. Their precision, reorientation and contact-recovery demands provide a stronger challenge than the retired 1,318-frame push_T reference. The decision uses observed manipulation and source alignment, not failure alone or elapsed time alone.
Retain provisionally
Retain the three-tube insertion task as the second additional contact task. The measured attempt succeeded after 5,117 native frames, with all three tubes upright and deeply seated.
The agent reoriented initially horizontal capped tubes, corrected the grasp-induced tilt and offset, selected fitting rack holes, and inserted them in stages. Final depths are 54.753, 54.850 and 56.392 mm against a requirement greater than 45 mm; upright errors are 1.683, 2.100 and 7.227 degrees against a requirement below 30 degrees. All three centers lie inside the rack footprint above its bottom. Both arms entered the native home region.
A reusable cap/tip localization and grasp-offset routine, reachable reorientation waypoints, staged descent and controlled gripper opening could transfer across saved layouts. The agent saved these routines with a warning to relocalize rather than reuse seed-0 coordinates. Transfer and efficiency gains remain hypotheses.
Instruction alignment: The instruction explicitly requires all three capped tubes, pointed ends down, capped ends up, upright deep seating and arm return. These match the unchanged native containment, depth, local-axis and home checks. All 165 shipped X5 layouts use the same capped conical tube asset and rack. No particular rack-hole identity or insertion order is imposed.
Measured instruction: Insert all three tubes upright and deeply into the rack, with their pointed ends down and their capped ends up. Leave the tubes seated in the rack, then return both arms to their starting poses.
Retain provisionally
Retain the ten-piece, four-peg assembly task. The measured attempt succeeded after 9,297 native frames with all four groups geometrically complete.
Original wrist observations show tilted and dropped star pieces followed by recovery attempts. The final native state has all ten pieces on the assigned pegs: group-anchor XY errors are 0.292, 0.903, 0.300 and 0.523 mm; the largest checked adjacent-piece XY error is 0.507 mm, all below the 1 mm native thresholds. Every piece also meets the native depth condition. Both arms entered the native home region.
Reusable wrist-based grasp-offset estimation, staged insertion, vertical retreat before lateral travel, and recovery by regrasping a tilted piece could reduce failures and work on later layouts. The saved memo explicitly tells future runs to reobserve positions. No cross-seed improvement is measured here.
Instruction alignment: The clarification resolves which pegs are correct by specifying the 4/3/2/1 shape groups in descending peg-height order, precise centering, lowering into each stack, and arm return. All 165 shipped X5 layouts have this same grouping and height order. Native success predicates and tolerances are unchanged.
Measured instruction: Stack all ten toy pieces on the four pegs, grouping pieces of the same shape together. Put the group of four on the tallest peg, the group of three on the next tallest, the group of two on the next, and the single piece on the shortest peg. Center each piece precisely on its peg and lower it into the stack. Leave all pieces on the toy and return both arms to their starting poses.
Native arm return is an end-effector tolerance of 15 cm on each position axis and 20 degrees in orientation. stop_on_success ends the episode as soon as this region is entered; exact home joints need not be reached. The tube run ended with a maximum positional-axis error of 147.84 mm, and the toy run with an 18.45-degree left-arm orientation error, both within those unchanged limits. Object placement conditions passed independently.
One seed-0 attempt per task; this is not a cross-seed reliability estimate. Modified instructions and unchanged native predicates are recorded separately. Video receipts state their exact replay validation scope.
robodojo/deposit_coin
Pick up the coin from its holder and deposit it through the slot of the coin bank. Let the coin fall fully inside the bank, then return both arms to their starting poses.
Complete lower-bank containment, with no bottom-contact or resting requirement; both arms return within native tolerances. This wording was revised after the retained historical run; that run remains labeled with its actual input.
robodojo/fasten_screws
Fit all three nuts onto the upright bolts of the same color. Leave each nut upright, centered on its bolt, and fully seated. Release all nuts, open both grippers, and return both arms to their starting poses.
Three same-color pairs; upright nuts precisely centered and deeply seated, all grippers open, arms returned. The task measures seating geometry, not tightening torque. This wording was revised after the retained historical run; that run remains labeled with its actual input.
push_T succeeded after 1,318 native frames. It is no longer in the selected task set. Its original evidence is preserved.
Historical push_T result