Robot policy gains are coming from action interfaces and execution constraints

Several papers treat the action interface as the main place to improve manipulation. ActionMap replaces pointwise action prediction with voxel heatmaps for translation, rotation, and gripper state. With the rest of OpenVLA-OFT unchanged, it raises the LIBERO four-suite average from 89.1% to 97.3%, and improves real Franka trials from 7/30 to 20/30 at full data.

LARA attacks the same control bottleneck through latent actions. It trains a Latent Action Model (LAM) together with a diffusion VLA policy, aligning the LAM latent with an intermediate policy feature. In the OXE-constrained LIBERO setting, LARA full reports 88.6 average success, above OpenVLA at 76.5 and its DiT-only version at 84.4.

Spline Policy is more methodological. It replaces fixed action chunks with spline parameters that can be resampled, constrained, and passed to controllers. The available excerpt gives no success-rate table, so its evidence is about compatibility and execution structure, not measured gains.

WIZARD focuses on the cost of adapting a frozen VLA policy to new tasks. It predicts task-specific Low-Rank Adaptation (LoRA) weights from a language instruction and a short demonstration video, then runs the adapted policy with no gradient update at test time.

The strongest result is on held-out LIBERO-Spatial: WIZARD reaches 0.40 average success, compared with 0.19 for MT-VLA with π0.5 and 0.02 for nearest-neighbor adapter retrieval. The paper also shows the boundary of the method. LIBERO-Object remains low at 0.03 average success, and full-task zero-shot completion on LIBERO-10 is still 0.00 in the excerpt.

Coarse-to-Control adds an internal plan, but the plan is made of action tokens. The model predicts coarse future action tokens first, then executable tokens conditioned on that plan. This keeps planning tied to wrist motion, gripper state, and waypoint structure. It reports 97.9% overall success on LIBERO and 83.3% on SimplerEnv-WidowX, with a 62.5% average in real-world tests using 50 demonstrations per task.

FLIGHT brings the same concern to unmanned aerial vehicles. The benchmark uses natural-language missions with continuous 6-DoF control, and its model splits work between a slower video-language module and a faster diffusion action model. The dataset contains 6,689 fine-grained navigation trajectories and samples continuous actions at 10 Hz.

STRIPS-WM covers high-level planning from images. It learns symbolic preconditions and effects from RGB action transitions, then plans in predicate space. Its numbers are about abstraction quality: in BlocksWorld it recovers 16 learned graph states for 16 ground-truth states with zero transition and applicability slack.

RhinoVLA is the clearest deployment paper in the set. It targets closed-loop manipulation on edge hardware, using Qwen3-VL to reduce visual-token cost and a continuous Action Expert for action chunks. The system reports 11.69 Hz end-to-end inference on the Huixi R1 edge system, above its 10 Hz control target.

The paper also names the runtime bottleneck it is trying to cut. In its π0.5 latency breakdown on Jetson AGX Orin, the VLM backbone and Action Expert account for more than 90% of runtime, and VLM MLP projections account for about 74.7% of VLM latency. Its cross-robot design uses a View Registry, a shared 72D state-action slot space, masks for missing robot dimensions, and robot-instance LoRA modules.