Robot policy papers concentrate on prediction, geometry, and harder VLA evaluation

Several papers treat prediction as the missing control layer for robot policies. AHEAD wraps a frozen 7B OpenVLA model with a 4.9M-parameter latent world model that forecasts task-relevant future visual tokens. Its reported gains are largest when object motion creates a timing problem: 79% to 97% success across 20 dynamic simulation scenarios, and 19/30 projectile catches on a physical xArm 7 where every listed baseline scores 0/30.

RoboDream uses generation for a different bottleneck: demonstration supply. It anchors video diffusion to rendered robot-only motion, then adds objects and scenes. Mixed real and generated training data reaches 62.5% average real-world success across four manipulation tasks, compared with 36.3% for the Real-50 baseline. Scaling mixed data reaches 72.5% to 73.75% in the reported setting.

Dexterity-BEV and Lie Diffuser Actor both make pose geometry explicit in the policy interface. Dexterity-BEV maps pixels into a shared 3D bird’s-eye-view (BEV) coordinate frame and expresses visual inputs, proprioception, and actions in that frame. It reports 89.9% average success on modified LIBERO camera and pose tests, while X-VLA and the 2D ablation are reported below 10%.

Lie Diffuser Actor keeps diffusion pose generation on SE(3), the rigid-body pose group, by adding noise in tangent space and mapping samples back with the exponential map. On CALVIN ABCD→D, it raises average task length to 3.584 versus 3.288 for 3D Diffuser Actor, and improves OpenVLA-OFT LIBERO Long success from 92.20% to 94.13% in a cross-architecture test.

RoboSemanticBench separates grasping from semantic target choice. It turns math, commonsense, and factual questions into pick-and-place episodes. Across 500 simulation episodes per model and suite, pi0.5 has the best average task success rate at 21.8%, yet its normalized Semantic Grounding score is only 5.2%. Several evaluated models have negative scores, meaning their target choice after a successful grasp falls below random-choice normalization.

FATE-VLA searches for failure-prone manipulation scenes using adaptive test generation. On GR00T-N1.6, the best variant raises the discovered failure rate to 65.3%, compared with 35.6% under random testing. On EO-1, failure discovery reaches 60.0%, compared with 36.7% under random testing. These results make the evaluation target more specific: find clustered failures across object, pose, and instruction conditions.

Two papers focus on making policy improvement less wasteful. EG-GRPO trains an aerial VLA policy with online rollouts plus one rule-based expert trajectory in each GRPO group. Starting from OpenVLA-OFT, it raises overall success rate from 26.1% under supervised fine-tuning to 55.6%, with intent-alignment score rising from 4.50 to 7.24. Its parallel rollout system cuts per-step rollout time by 43.5%.

CSIL++ improves pi-0.5 on sparse simulated manipulation tasks with a learned coherent reward derived from demonstrations. The ensemble variant reaches at least 90% success on five of six tasks, including Threading at 0.92 versus 0.14 for the pi-0.5 baseline. The evidence is simulation-only in the available excerpt, but the gains show a clear route for improving large behavior models with fewer hand-written rewards.