A single robot-control policy reported strong results across simulated manipulation benchmarks and controlled tasks on a Piper robot, yet its performance fell sharply when it was moved directly to a different robot body. UCAG-P reached 98.3% on LIBERO, 88.66% on RoboTwin Easy, 89.20% on RoboTwin Hard and 62.0% on RoboCasa GR-1. In a direct ALOHA-to-ARX replacement, zero-shot success was 35.0%.
The result addresses the paper's central question: can one UCAG-P checkpoint operate across heterogeneous embodiments—different robot bodies and control interfaces—remain robust when test conditions shift, and transfer to new physical robots?
Different bodies, one shared target
The policy was trained on an embodied corpus of 1,020,672 episodes lasting 6,373.586 hours across eleven subsets. The material spanned single-arm, dual-arm, humanoid and human-hand embodiments.
That total was unevenly distributed. Simulation supplied 3,767.511 hours, compared with 266.348 hours of real-robot data and 2,339.727 hours of human-hand video. Robot and simulation data together made up 63.29% of the hours, while human video accounted for 36.71%.
UCAG-P's proposed common interface is a camera-centric action space. It describes movement through two anchors: the wrist or end effector, and the gripper or grasp center, rather than through each embodiment's own controls. In practical terms, the shared target is expressed around what the camera observes and the manipulation points, while the robot-specific control interface is handled later.
The later translation step combines the shared motion with the transform from camera to robot base, a Jacobian map of how a robot's joints relate to end-effector movement, and the robot's current state. It then produces embodiment-specific executable actions.
Training was progressive. One stage focused on shared visual-geometric prediction, another on geometry-conditioned translation, and a final stage jointly optimized the camera-centric targets and robot-control outputs. Missing labels were masked, so each sample contributed only the targets it actually contained.
Strong tests, uneven robustness
Simulation used the same final checkpoint without benchmark-specific fine-tuning. For the Piper comparison, each task used 100 demonstrations and 20 closed-loop trials under matched conditions.
On the standard LIBERO suites, the aggregate success rate was 98.3%. The suite scores were 98.8% for LIBERO-Spatial, 98.6% for LIBERO-Object, 99.2% for LIBERO-Goal and 96.4% for LIBERO-Long. The model also reported 88.66% on RoboTwin Easy, 89.20% on RoboTwin Hard and 62.0% on RoboCasa GR-1.
Under distribution shift, the picture was less uniform. In the zero-shot LIBERO-Plus test, the total score was 82.0%. Robot-state, lighting and background changes were more favorable, while camera changes, sensor noise and object-layout shifts were harder.
On the Piper tasks, UCAG-P reported 60% success for bread grasping, 90% for drawer opening and 75% for bowl stacking. The matched π0.5 comparison reported 20%, 85% and 65%, respectively, so UCAG-P's observed rate was higher on each listed task.
The transfer gap remains
Transfer across robot bodies was the harder test. In the direct ALOHA-to-ARX replacement, the model retained nonzero success at 35.0%, but that was substantially below performance on the source embodiment. The authors identify persistent morphology and kinematic mismatch as a major limitation of this transfer.
One reported comparison examined the model's output module. Without vision-language-model features, it recorded 58.3% success; with attend-pooled VLM features, it recorded 62.0%, a difference of 3.7 percentage points. That is an observed result from one ablation, not evidence that the feature choice will have the same effect elsewhere.
What the results do not settle
The study reports observed percentages but no confidence intervals, inferential tests or formal power rationale. The gaps should therefore be read as results from the reported evaluations, with their repeatability and statistical uncertainty left open.
The authors also flag error propagation from geometric estimation or calibration, persistent morphology and kinematic mismatch during transfer, and the limited scale of the controlled real-world evaluation. The evidence is confined to computational training, benchmark simulations and controlled Piper tasks; it does not establish broad deployment robustness.
Paper data and sources
Original title: One Policy, Many Embodiments: Unified Camera-Centric Action Geometry Pre-training for Heterogeneous Embodied Manipulation
Authors: Xiaomi Embodied Intelligence Team, University of Macau, : et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text