Robot research treats 3D contact and real testbeds as the main evidence

Several robot papers put geometry closer to the action output. GaussianDream trains a VLA policy with 3D Gaussian reconstruction and short-horizon scene-flow prediction, then removes the decoding heads at inference and keeps only learned prefix tokens. It reports 98.4% average success on LIBERO and 52.6% on RoboCasa Human-50.

PointACT makes the action decoder attend to multi-scale point-cloud features. Its reported LIBERO average is 96.0%, with a 17.9-point gain over SpatialVLA in the same table. DISC tackles a different grounding failure: the instruction generates a task-specific policy, while the control policy sees only observations. That design reaches 94.3% on LIBERO-90 and 86.4% on a real 9-task shared-context setup.

Evaluation work is becoming more concrete and cheaper to repeat. VLA-REPLICA defines a real-world VLA testbed built from about $1050 of parts, with 10 manipulation tasks, 500 demonstrations, and 90 test scenes. In its in-distribution results, π0.5 leads the tested policies at 0.54 average success, which also shows how much room remains on low-cost physical setups.

stable-worldmodel addresses the world-model side with shared data loading, baseline training, model-predictive control planning, and evaluation across robotics and control tasks. Its Lance data layer reaches 4,815 Push-T samples per second locally, compared with 1,416 for HDF5. Lost in Fog adds a driving VLA stress test: Alpamayo R1’s chain-of-causation explanation changes correlate with much larger trajectory deviations under noise, lighting changes, and fog.

Dexterous-hand papers focus on control signals that can survive hardware limits. Proprioceptive Transformer uses only joint position and velocity histories on a tendon-driven ORCA hand. On a 55 mm cube, its direct-joint-sensing policy reaches 11.83 RPM with 100% rotation accuracy and no drops across three 60-second trials.

SafePBDS adds explicit safety constraints to geometric motion policies. It solves for configuration-space acceleration while enforcing task-manifold safety conditions as hard constraints. On a Franka Panda plus Allegro Hand, it reports 92.5% success across 120 household-object grasping trials and more than 360° yaw reorientation in both directions.

Demo-JEPA treats a demonstration as a goal signal that the target robot must realize with its own dynamics. A V-JEPA-style encoder maps observations to latent states, a Dreamer Predictor estimates a target-compatible future latent goal, and the robot plans actions toward that goal with its own world model.

The results are strongest in transfer settings. In simulation, Demo-JEPA averages 0.45 success on cross-embodiment bridging and 0.36 on zero-shot generalization, ahead of VPP and XSkill in the reported comparisons. In-domain behavior grounding is weaker, with 0.31 success in RLBench simulation and 0.43 in real-world behavior grounding.