VLA action interface evaluation

Teams training OpenVLA-OFT, π0.5, or similar VLA policies should test the action head as a replaceable component. ActionMap keeps the backbone fixed and replaces single-point action prediction with voxel heatmaps for translation, rotation, and gripper state. The training target is a soft Gaussian blob over the action grid, and inference recovers a continuous command with top-k soft argmax.

The practical test is small: run the same training budget with the native decoder and with a voxel heatmap head on LIBERO-Spatial and LIBERO-Long, then repeat on one real Franka pick or place task with grasp-position error logged. ActionMap reports a LIBERO four-suite gain for OpenVLA-OFT from 89.1% to 97.3%, a 10% data LIBERO-Spatial gain from 67.2% to 93.2%, and real Franka full-data trials rising from 7/30 to 20/30. Those numbers make the decoder swap a credible first experiment for teams seeing millimeter-scale end-effector errors or low-data failures.

Robot teams trying to run closed-loop VLA control on onboard hardware should profile the VLM backbone, visual-token count, and Action Expert before changing the controller. RhinoVLA’s report gives a concrete latency target and breakdown: it reaches 11.69 Hz end-to-end inference on the Huixi R1 edge SoC, above a 10 Hz control target, while its π0.5 analysis on Jetson AGX Orin puts more than 90% of runtime in the VLM backbone and Action Expert.

A useful build is a deployment harness that records image-token count, VLM projection time, Action Expert time, and end-to-end command rate for each camera configuration. RhinoVLA’s design also points to the data-interface work needed for mixed robot fleets: a View Registry for camera role and modality, a shared 72D state-action slot space, masks for missing dimensions, and robot-instance LoRA modules. The report does not give detailed task-level scores in the excerpt, so the first adoption check should pair the latency target with a fixed manipulation success suite on the target robot.

Labs that already maintain task-specific LoRA adapters can test a generated-adapter workflow for new manipulation tasks with no target action labels. WIZARD trains expert LoRAs on known tasks, encodes a new task from its language instruction and short demonstration video, and predicts the LoRA weights for a frozen VLA policy in one forward pass. The robot then runs the adapted policy without gradient updates at test time.

The workflow fits task onboarding where a human can provide a short video but cannot provide synchronized robot actions. The first benchmark should separate spatial rearrangement tasks from object-identity changes. WIZARD reaches 0.40 average success on held-out LIBERO-Spatial versus 0.19 for MT-VLA with π0.5 and 0.02 for nearest-neighbor adapter retrieval, but LIBERO-Object remains low at 0.03 and full-task zero-shot completion on LIBERO-10 is 0.00 in the excerpt. That boundary matters for deployment planning: use the method first where task geometry changes more than the object distribution.