Robot papers target failure recovery, domain data, and action-conditioned prediction

Two papers attack execution drift directly. RePO-VLA trains on success, failure, and recovery rollouts with separate labels, then deploys with a fixed high value condition and no online failure detector. Its reported adversarial success rises from 20% to 75% on average, with FRBench-Sim covering 23,453 bimanual episodes across 46 tasks.

CAPS keeps the base VLA policy unchanged and spends extra inference only when uncertainty rises. It samples future action chunks with a power distribution and uses Metropolis-Hastings search when entropy crosses a threshold. On RoboTwin 1.0 with π0, average success is 47.4%, compared with 32.2% for π0 and 41.3% for π0 plus TACO. On Simpler-WindowX, it reports 60.5% average success, ahead of π0 and several VLA baselines.

SABER shows the data problem in retail robotics with concrete numbers. The dataset uses about 100 hours of grocery-store capture and converts human activity into 44.8K training samples: 25K latent-action sequences, 18.6K hand-pose trajectories, and 1.2K whole-body motion sequences. Post-training GR00T N1.6 on these streams reaches 29.3% mean success across 10 RoboBenchMart tasks, compared with 13.4% for simulation-only fine-tuning.

DRIS addresses a different data gap in dexterous sim-to-real transfer. Each simulated episode contains several object instances with varied dynamics, all advanced under the same robot action. The paper claims zero-shot real transfer for flat-plate reactive catching, but the available excerpt does not provide a real-trial success rate, so the evidence is stronger for the training design than for the deployment result.

World-model papers focus on prediction that can guide control. Sub-JEPA regularizes latent spaces through low-dimensional Gaussian tests instead of forcing the whole latent vector into an isotropic Gaussian. Against LeWorldModel, it improves planning success on all four reported tasks, including Two-Room at 95.00±2.76% and OGB-Cube at 76.33±5.99%.

DeformMaster tackles ropes, cloths, packages, and soft toys with a physics-neural model learned from real videos. It rolls material particles forward with differentiable MPM, adds bounded neural velocity correction, and renders new action rollouts with Gaussian Splatting. On 20 real PhysTwin sequences, it reports IoU 0.748, Chamfer 0.011, PSNR 25.41, and online interactive rollout above 15 fps.