Simulation and policy weights
QLM-Bench Standard100 uses ten MuJoCo tasks, ten held-out scenes per task, and seeds 42, 43 and 44 for 300 episodes per policy. Task time limits range from 24 to 60 seconds. The benchmark has 1,000 synthesized demonstrations. Stage 1.1 fine-tuned LingBot-VA on 800 of them and reached 77.0% with 25 video and 50 action sampling steps. Its Stage 1.2 OPSD student (rank-256 LoRA, 300 updates, EMA weights) reached 76.7% with one video and two action steps. Stage 2.1 uses separate rank-32 push-box (r14) and open-door (r04) RL teachers. Stage 2.2 starts from the Stage 1.2 student and distills the specialist teachers by task. The released Stage 2.2 M-OPD student (300 updates, EMA weights) reached 85.0% in simulation. Both students load the Stage 1.1 teacher as the inference base. LIBERO arm weights are supplementary experiments.
Physical trials
The authors ran the Stage 1.2 OPSD student trained on simulation data only on a Unitree Go2 Air, using the Go2 Air’s built-in egocentric camera and a RealSense D435i raised above the trunk for a top-down view. The robot carried an RDK X5 board. An off-board RTX 5070 Ti workstation ran the policy and VAE over wired Ethernet. RAMBO executed the base and front-left foot commands through whole-body control. Successes were 3/10 for button pressing, 1/10 for ball kicking and 3/10 for box pushing (7/30 total). On the physical robot, sampling took 0.159 s per action chunk. The complete control loop, including observation encoding, took 0.66 s. Each chunk covered 0.32 s of robot time, so the loop fell short of real time and the controller held its last command between chunks.
Real-robot deployment package
The authors provide a ROS 2 Jazzy package for Unitree Go2. The documented setup uses Ubuntu 24.04, Python 3.12 and an NVIDIA GPU on the host, wired Ethernet, and a RealSense D435i on an RDK X5 camera host alongside the Go2 front camera. The package uses RAMBO for whole-body control and QLM-Bench [live] for the qlm-live-v1 policy connection. The main ROS dependencies are rclpy, rcl_interfaces, unitree_go, unitree_api, geometry_msgs, std_msgs and sensor_msgs, with launch, launch_ros and ament_index_python at runtime. Fast DDS is the default, with Cyclone DDS documented as an option. The default policy is Stage 1.2 OPSD, which uses the Stage 1.1 teacher as its inference base. The paper reports hardware trials on Go2 Air. Physical results for Go2 Pro, X and EDU, as well as for the other policy options, are unconfirmed.
See sections 4–5, tables 1 and 7, and appendices E and G–H of the paper, as well as the authors’ Limitations and Future Work update.