The MetaFine Open Competition (RoboFineMani workshop, CoRL 2026) evaluates fine-grained manipulation beyond binary success. Its five simulated tasks, from grasping a bottle by an instructed part to inserting letter-shaped pegs with 4 mm clearance, are scored on perception under lighting and camera changes, instruction understanding, and behavior. This report presents Team Friday's solution, which got second place with an overall score of 0.947, more than twice the π0 baseline. Instead of fine-tuning a large vision-language-action (VLA) model, we train one 19M-parameter policy per task with a separate instruction classifier and focus on the data, all generated in the simulator by a privileged scripted teacher, including its corrections of the student's mistakes. By analyzing failed episodes, we find that a teacher that almost always succeeds can still teach a memoryless student badly: its pauses become stalls, its fixed habits become shortcuts, and it never shows the errors the student makes. Making the teacher imitable raised the official score on letter insertion, the hardest task, from 0.44 to 0.82 with 32% more data and the same architecture, and finally to 0.88. We summarize practical lessons on demonstration data and evaluation, which we hope will help in imitation learning for precise manipulation.
Figures & tables
Figure 1: Overview. Top: offline data generation with a privileged teacher and takeovers (Sec. 3 ). Bottom: the student policy (Sec. 2 ), one control network per task plus a shared instruction classifier. For T5, a second image stream sees a gripper crop at about twice the resolution.
Ours (Team Friday)
π0 baseline
Task
Episodes
P
U
B
Score
Episodes
P
U
B
Score
T1 grasp part
1,410
0.865
0.900
0.900
0.888
200
0.296
0.35
0.35
0.332
T2 move mug
1,192
0.983
0.983
0.983
0.983
300
0.892
0.95
0.95
0.931
T3 toggle switch
1,955
1.000
1.000
1.000
1.000
200
0.346
0.60
0.60
0.515
T4 cubes into boxes
896
0.995
0.983
0.983
0.987
300
0.192
0.35
0.45
0.331
T5 insert letter
9,674
0.880
0.875
0.875
0.877
400
0.025
0.05
0.05
0.042
Table 1: Official results (task score: mean of P, U, B; overall: mean over tasks). Episodes: training episodes per task (ours: teacher demonstrations and takeovers; π0 : planner demonstrations).
Network
Episodes
Change in the data
Local
Official
t5_a
1,541
plain teacher demonstrations
13/80
–
t5_c
4,067
off-centre grasps, seating from the true in-hand pose, staged misalignments; takeovers
34/80
0.442
t5_f
3,818
settles capped at 3 steps, release once the peg is home, descent noise; takeovers
61/80
–
t5_g
5,358
takeovers resume from the current height instead of rising first
71/80
0.817
t5_j
9,674
release only once seated; two more takeover rounds
–
0.875
Table 2: T5 development. Local: successes in 80 clean episodes (same seeds in every row). Official: leaderboard T5 score. The network is unchanged from t5_c on (t5_a: centre instead of gripper crop, fewer auxiliary targets); t5_j was fine-tuned from t5_g, the others trained from scratch.
Fine-grained manipulation couples instruction grounding, local visual precision, and sustained execution. We present FineCraft, a framework that adapts pretrained models for fine-grained manipulation in the MetaFine Challenge. A task-routed mixture of policies selects independently adapted DINOv2-DiT or pretrained π0.5 policies through a common observation-action interface. Local visual features help identify the instructed object parts, demonstrations and expert continuations teach the required actions, and new observations guide replanning during execution. FineCraft ranks fourth with an overall score of 0.878, exceeding the published reference on all five tasks. Official results and local diagnostics characterize performance across visual perturbations, instruction changes, and multi-stage manipulation. Code is available at https://github.com/eqcy/FineCraft
Shuaike Zhang, Ziheng Quan, Junhao Ran +3
Harbin Institute of Technology, Shenzhen · Xi’an Jiaotong University · Shenzhen Loop Area Institute HITSZ Robotics
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.
Lifan Zheng, Guotao Chen, Zhenyang Li +3
Southeast University · The Hong Kong Polytechnic University · Monash University +1
Fine-grained robotic manipulation requires instruction grounding, robust perception, and precise contact control. We describe our submission to the MetaFine Open Competition at CoRL 2026. Five independently trained models use a common architecture with frozen DINOv2 and CLIP encoders, trainable language-conditioned feature modulation, and task-specific action heads. We improve demonstrations through contact-aware grasp filtering, instruction-matched data, and expert correction of insertion failures. At inference, task-dependent strategies include photometric test-time augmentation, stochastic action averaging, and uncertainty-triggered replanning. Our submission achieves an official overall score of 0.914 and ranks third among 15 teams. We analyze the models and their task-specific behavior.
Chanho Kim, Soyun Choi
Department of Immersive Media Engineering Sungkyunkwan University South Korea