Robot VLA claims now need real control evidence

Several papers treat VLA deployment as a post-training problem with measurable robot time and failure recovery. EXPO-FT reports 30/30 success on eight real manipulation tasks after an average of 19.1 minutes of online robot data. BORA adds an offline critic and a small human-guided residual actor for dexterous hand control, reaching 86.0% average success across five Franka arm plus 12-DoF hand tasks. A continual-learning study gives the counterweight: plain sequential fine-tuning can erase earlier skills, while experience replay with fixed action normalization raises the final average score to 93.5.

The strongest control papers add structure around action prediction. OASIS predicts future end-effector poses in SE(3) before decoding action chunks, and reports 97.6% average success on LIBERO plus 89.2% average success in real-world tests. Qwen-VLA uses embodiment-aware prompts and a shared action-and-trajectory format to cover manipulation, navigation, and trajectory prediction with one policy. VLA-Pro stores task-specific LoRA adapters as procedural memories and retrieves them at inference, raising real-world success on held-out UR7e tasks from 5.8% to 65.0%.

Contact quality becomes part of the evaluation target. Tabero adds tactile tokens and closed-loop force commands, then measures success together with grip and applied force. It reports over 70% lower average grip force under gentle instructions. CoP compresses tactile taxel readings into contact force and contact location for sim-to-real dexterous manipulation, reaching 0.78 real peg-in-hole success across six shapes. Mag-VLA extends the same execution focus to microscope-scale bimanual magnetic manipulation, with 90% approach success and lower transport success as path curvature rises.

Cost is a visible constraint in the week’s evidence. ProgVLA uses a 0.1B-parameter policy with compressed multimodal tokens and progress heads, reaching 91.1% average success on LIBERO and 88.6% on LIBERO Long. Its real PiPER-arm tests are more modest at 68% success over 100 trials, which gives a useful check against benchmark-only claims. The day-level trend also groups HyperSim and SDPG with efforts to reduce real-data or GPU cost, placing efficiency alongside execution quality.