Robot learning work centers on deployable control and measurable failure modes

VLA manipulation papers put more weight on real execution metrics. EXPO-FT keeps a pretrained π0.5 policy and trains a lightweight edit policy with off-policy reinforcement learning. It reports 30/30 final successes on each of 8 real-world manipulation tasks after an average of 19.1 minutes of online robot data.

OASIS attacks the action decoding problem directly. It predicts an 8-step SE(3) end-effector trajectory, meaning 3D position and rotation, before producing 6-DoF actions and gripper commands. The paper reports 97.6% average success on LIBERO and 89.2% average success in real-world tests on Franka Research 3 and Kinova Gen3 robots.

The deployment evidence is more operational than architectural. The Siemens factory-floor case study fine-tunes Pi0.5 for transparent accessory-bag packaging, using 2,535 episodes and about 10 hours of data. Its failure breakdown is concrete: in unconstrained trials, bag contents left above the product account for 65% of failed episodes, with multiple-bag grasps at 23% and poor or failed grasps at 15%.

A separate VLA reliability paper gives an information-theoretic bound for capability and adversarial reliability. It cites OpenVLA-7B dropping from above 95% LIBERO success to under 5% under a 16/255 PGD image attack, then validates the bound with zero violations across 320 cells covering Gaussian proxies, OpenVLA, LIBERO suites, attack types, horizons up to 10, and two action-head designs.

World-model papers focus on state representations that support planning and policy improvement. TC-WM compresses frozen visual foundation features and proprioception into task-centric latents, then trains action-conditioned dynamics for planning or reinforcement learning. The reported coverage spans 9 benchmarks across Robomimic, D4RL, navigation, locomotion, manipulation, and simulated plus real-world settings, though the summary does not include exact score tables.

MBDPO trains a diffusion action policy inside a latent world model, so imagined rollouts guide the same policy distribution used at execution time. Its strongest numeric claims are scale and breadth: model size grows from 1.7M to 340M parameters, online experiments cover 4 benchmark suites and 121 tasks, and an 8-task drift study reports lower action drift than TD-MPC2.

LeJEPA adds theory for when a learned representation is a usable world model. Under Gaussian latent variables and stationary additive-noise positive pairs, the optimum recovers the latent state up to an orthogonal transform. The paper also links that recovery to finite-horizon planning when costs are invariant under such transforms.