A high-frequency mid-level layer that picks which whole-body policy runs, and what goal it gets.
Weirui Ye1*, Yen-Jen Wang2,5*, Elvis Hsieh2*, Zhiyang Dou1, Shibo Zhao3,5,
Pieter Abbeel2,5, Yilun Du4, Leslie Pack Kaelbling1, Tomás Lozano-Pérez1
1MIT CSAIL 2UC Berkeley 3CMU 4Harvard 5Amazon Frontier AI & Robotics *Equal contribution
Planners operate on task timescales; whether their command is feasible changes on control timescales. EPO sits in between. For every frozen goal-conditioned policy it learns one energy: will this goal stay feasible, and will it hit the target? At 50 Hz it runs the lowest-energy policy–goal pair, and at 10 Hz it refines each policy's goal down its own energy. The planner and the policies stay unchanged.
This repository contains, for the Unitree G1 in Isaac Lab:
- Four goal-conditioned whole-body control policies, which you can also drive directly from your own planner or by teleoperation (keyboard / VR).
- Their energy models, and the EPO orchestrator that runs them as one controller.
- Code to train an energy for your own policy (~30 GPU-minutes); nothing else is retrained.
- The evaluation protocol that reproduces the paper's numbers.
Videos and results: project page.
| Policy | Goal space | Trained with | Policy | Energy |
|---|---|---|---|---|
gen |
velocity, both hands, height (16-D) | pushes ±1 m/s | checkpoints/policies/gen |
checkpoints/energy/gen.pt |
gen_robust |
16-D | pushes ±2 m/s, 0–2 kg wrist payloads | checkpoints/policies/gen_robust |
checkpoints/energy/gen_robust.pt |
stand |
hands, height (13-D) | pushes ±1 m/s | checkpoints/policies/stand |
checkpoints/energy/stand.pt |
stand_robust |
13-D | pushes ±2 m/s, 0–2 kg wrist payloads | checkpoints/policies/stand_robust |
checkpoints/energy/stand_robust.pt |
All four are recurrent (LSTM) policies at 50 Hz that share one observation and one 16-D command:
[vx, vy, wz, left hand xyz, right hand xyz, left hand rot, right hand rot, height].
Details: docs/policies.md.
Linux, an NVIDIA GPU, Isaac Sim 5.0, Isaac Lab 2.2.0, Python 3.11. Install Isaac Lab following its guide, then in that environment:
git clone https://github.com/YeWR/EPO.git && cd EPO
pip install -e .Run a released policy (installation check):
python legged_lab/scripts/run_policy.py --checkpoint checkpoints/policies/gen_robust/model.pt \
--num_envs 4 --steps 200 --headlessRun EPO with all four policies on flat ground:
python scripts/evaluate.py --library M4 --terrain plane --out results/epo_m4
python scripts/analyze.py results/epo_m4Use EPO in your own control loop:
from epo.energy import load_energy
from epo.goals import LIBRARIES, POLICIES
from epo.orchestrator import EPOConfig, Orchestrator
library = {k: load_energy(f"checkpoints/energy/{k}.pt", POLICIES[k], device) for k in LIBRARIES["M4"]}
orch = Orchestrator(library, num_envs, EPOConfig())
# every control tick (50 Hz)
running = orch.step(t, state, height_scan, z_task, mask) # which policy each robot runs
goal = orch.goals[running[0]] # that policy's refined goalSee epo/rollout.py for the full loop in Isaac Lab.
- Put your RSL-RL checkpoint at
checkpoints/policies/<name>/model.pt(with itsparams/). - Register it in
epo/goals.py: aPolicySpecinPOLICIES(its goal dimensions), and add it to a library inLIBRARIES. Add<name>to the--policychoices inepo/train_energy_args.py. - Train its feasibility model F and outcome model T, and pack them into one energy:
python scripts/train_energy.py --policy <name> --model F --out runs/<name>_F.pt
python scripts/train_energy.py --policy <name> --model T --out runs/<name>_T.pt
python scripts/pack_energy.py --policy <name> --F runs/<name>_F.pt --T runs/<name>_T.pt \
--out checkpoints/energy/<name>.ptThe existing policies and energies are untouched. Details: docs/energy.md.
The evaluation protocol, ablation flags and expected numbers are in docs/evaluation.md. The code reproduces the paper's runs tick for tick on the same GPU setup.
| Path | What it is |
|---|---|
epo/ |
The method: goal space, energy models, goal refinement, policy selection, orchestrator |
scripts/ |
evaluate.py, analyze.py, train_energy.py, pack_energy.py |
checkpoints/ |
Released policies and energy models |
legged_lab/ |
The G1 Isaac Lab task the policies were trained in, with train.py, play.py (keyboard / VR teleop), export_onnx.py |
docs/ |
Policies · Energy models · Evaluation |
@inproceedings{epo2026,
title = {Energy-Based Policy Orchestration},
author = {Ye, Weirui and Wang, Yen-Jen and Hsieh, Elvis and Dou, Zhiyang and Zhao, Shibo and Abbeel, Pieter and Du, Yilun and Kaelbling, Leslie Pack and Lozano-P{\'e}rez, Tom{\'a}s},
booktitle = {Conference on Robot Learning (CoRL)},
year = {2026}
}BSD-3-Clause; see LICENSE.txt. The environment is derived from Legged Lab and Isaac Lab
(BSD-3-Clause). legged_lab/utils/xr/webxr_bridge.py contains coordinate conventions from Unitree's televuer
(MIT; notice in the file).