WAM-Policy: World Action Models with On-Policy Distillation for Quadrupedal Loco-Manipulation
Paper available
Authors & affiliations (6)
Author order: Original order
- First published
- 2026-11-18
- Latest paper revision
- v1 · 2026-11-18
- Venue
- Not provided
WAM-Policy adapts LingBot-VA into a language-conditioned policy for quadrupeds that manipulate with the front-left foot. The authors evaluated cross-morphology transfer, on-policy self-distillation and multi-teacher distillation in MuJoCo, and tested the Stage 1.2 OPSD student on a Unitree Go2 Air.
View full model descriptionTo run QLM-Bench, check out the qlm-bench branch separately and use a dedicated simulator environment with Python 3.11 or later. Install the QLM-Bench [live] extra in the WAM-Policy environment to connect the policy and simulator via the local qlm-live-v1 protocol. Set QLM_BENCH_SRC and QLM_PYTHON in configs/site.env to the simulator source directory and interpreter.
Paper available
Author order: Original order
LIBERO arm evaluation and ablation guide
SourceVerified on · 2026-11-20
Verified on · 2026-11-18
Some training datasets are public; this does not establish access to the complete training mixture.
| File | Access | Weights license | Commercial use |
|---|---|---|---|
| Stage 1.1 transfer teacher Training data: Public access | Access | Weights license | Commercial use |
| Stage 1.2 OPSD student S300 EMA Training data: Public access | Access | Weights license | Commercial use |
| Stage 2.1 push-box specialist Training data: Public access | Access | Weights license | Commercial use |
| Stage 2.1 door-opening specialist Training data: Public access | Access | Weights license | Commercial use |
| Stage 2.2 M-OPD student S300 EMA Training data: Public access | Access | Weights license | Commercial use |
| Supplementary LIBERO OPSD S300 EMA | Access | Weights license | Commercial use |
| Quadruped ablation · Teacher cache only S300 EMA Training data: Public access | Access | Weights license | Commercial use |
| Quadruped ablation · Teacher cache only S2000 EMA Training data: Public access | Access | Weights license | Commercial use |
| Quadruped ablation · Streaming only S300 EMA Training data: Public access | Access | Weights license | Commercial use |
| Quadruped ablation · OPSD S900 EMA Training data: Public access | Access | Weights license | Commercial use |
| LIBERO arm ablation · Teacher cache only S300 EMA | Access | Weights license | Commercial use |
| LIBERO arm ablation · Streaming only S300 EMA | Access | Weights license | Commercial use |
| LIBERO arm ablation · Collected once S300 EMA | Access | Weights license | Commercial use |
| LIBERO arm ablation · SFT only S692 EMA | Access | Weights license | Commercial use |
| LIBERO arm ablation · Teacher cache only S2000 EMA | Access | Weights license | Commercial use |
| LIBERO arm ablation · OPSD S900 EMA | Access | Weights license | Commercial use |
| Quadruped ablation · RGB + proprioception S5000 Training data: Public access | Access | Weights license | Commercial use |
SourceVerified on · 2026-11-20
Physical results are reported by the authors for the Stage 1.2 OPSD student. The Stage 2.2 aggregate is a simulation result. Arm checkpoints are LIBERO experiments. The deploy/ directory uses CC BY-NC 4.0. The remaining WAM-Policy main-branch code and the officially released policy weights use Apache-2.0. The QLM-Bench simulator branch code, the QLM-Bench dataset, and the RAMBO controller weights distributed with it use CC BY-NC 4.0. The shared command and live interface files on main are provided under that branch’s Apache-2.0 license. The quadruped evaluation used one quadruped platform, one whole-body controller and one manipulating leg (the front left). Standard100 varied the scenes of the ten training tasks. It did not test transfer to other legged robots or generalization to unseen tasks or instructions. Stage 2.1 tuned only box pushing and door opening. The policy was trained entirely in simulation and performed worse on the physical robot. For v2, the authors plan to replace the two cameras with one wide-angle fisheye camera near the egocentric camera position, paired with LiDAR. The announcement gives a planned sensor setup. Release and weight availability dates are still unannounced. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
Apache-2.0
Cross embodiment
Official
Official weights
Actions / Video
Future video and action sequences
Quadruped / Robotic arm
Multiple tasks
Reinforcement learning / Supervised fine-tuning
Derived from: LingBot-VA · LingBot-VA base ↗
MuJoCo benchmark and data for ten quadrupedal loco-manipulation tasks.
800 episodes · CC BY-NC 4.0
No verified independent benchmark results are recorded for this version or the selected weights.
Publisher evaluation · Reported; not independently reproduced
QLM-Bench · Standard100 ↗Success rate: 85 %
mujoco-standard-test100-v1; seeds 42, 43, 44
72 episodes on the RTX 5070 Ti workstation with CPU VAE, and 228 on a second workstation with GPU VAE and concurrent clients.
MuJoCo Standard100, 100 cases across 10 tasks, 3 seeds, 300 episodes. Author-reported aggregate. Video/action steps 1 / 2.
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.