Robot learning papers put VLA policies under execution pressure

RoboWorld gives the cleanest world-model result. It evaluates eight open robot policies through 4,186 generated rollouts, then matches the RoboArena real-world ranking with Pearson r=0.989 and Spearman rho=0.970. The key design is closed-loop rollout generation plus a 0–5 task-progress vision-language model judge, which is more informative than binary success scoring in the reported ablation.

Other papers make the control link explicit. ABot-M0.5 defines a world action model (WAM) for mobile manipulation that predicts future video, latent motion, and executable robot actions in sequence. The tutorial paper also tries to tighten terminology by defining robot world models as action-conditioned predictors and grouping WAM designs by how predicted futures connect to actions. The measured evidence is strongest for RoboWorld; ABot-M0.5 and the tutorial mainly clarify model design choices in the available excerpts.

FurnitureVLA is the main long-horizon manipulation result. It decomposes IKEA-style bimanual assembly into language-conditioned subtasks and predicts continuous subtask progress alongside a 14-dimensional dual-arm action. In simulation, average full-assembly success rises from 0.48 for monolithic finetuning to 0.80 across LACK, KALLAX, and IVAR.

Safety work adds another execution test. The constrained flow-matching paper edits predicted 10-step end-effector trajectories during denoising, using control barrier function constraints before the action chunk is finalized. On SafeLIBERO, it reports 82.81% collision avoidance and 81.62% task success, compared with 18.69% and 50.88% for unguided π0.5. The gain comes with slower execution, so the result is a safety-throughput tradeoff rather than a free improvement.

Tactile pretraining receives a concrete data contribution. H-Tac contains 160 hours, more than 300 tasks, and over 135,000 episodes with tactile, action, vision, and language signals. Transferable Tactile Pre-Training adds a tactile expert that predicts future touch readings alongside action chunks. The excerpt gives dataset scale, while downstream success rates are not available in the supplied text.

Data access also gets practical attention. The Daft post shows per-frame search over Apple EgoDex using SigLIP-2 text-video embeddings plus hand-pose geometry. Its examples retrieve events such as stapler grips, block lifts, and t-shirt folding clips. That matters for robotics teams because fine manipulation failures often sit inside long unlabeled episodes.

Certified sensing clocks add a third sensor-facing thread. A frozen 3D VN-JEPA world model exposes a drift-aware deadline for when an agent should re-sense, with held-out interval certificate violation upper bounds below the stated 0.15 target in three tests.

Adaptation and serving papers focus on what happens after a policy leaves its training setup. DART adapts a VLA to a new camera, lighting, sensor, or embodiment condition using one target-domain demonstration for one task plus a matching source demonstration. It extracts a domain vector from weight updates and adds it to the base policy, but the supplied excerpt does not include success-rate tables.

BIFROST addresses sim-to-real transfer through a shared latent history encoder trained on paired cross-domain segments. In sim-to-sim navigation, it reports 0.68 ± 0.08 success for top-down views and 0.50 ± 0.08 for egocentric views, with weaker direct-transfer baselines at 0.19 ± 0.04 and 0.03 ± 0.02. The excerpt claims sim-to-real tests, but the quantitative tables are absent.

ROSA treats robot foundation model inference as a fleet scheduling problem. On eight H200 GPUs and up to 64 virtual robots, it reports up to 12.06× higher SLO-qualified factory productivity than dedicated serving baselines, and up to 2.44× over shared-server baselines without its scheduler.