[CoRL 2026] RPL: Learning Robust Humanoid Perceptive Locomotion on Challenging Terrains

Yuanhang Zhang, Younggyo Seo, Juyue Chen, Yifu Yuan, Koushil Sreenath, Pieter Abbeel^†, Carmelo Sferrazza^†, Karen Liu^†, Rocky Duan^†, Guanya Shi^†

Project Website | Paper | Video

Note

Research release only. For potential issues and future support, please open an issue here or directly contact Yuanhang Zhang (yuanhang0610@gmail.com). Any integration of RPL features into Holosoma mainline is a separate effort and will not be reflected here. We use IsaacGym as the training backend since IsaacSim is not fully supported yet (force perturbation not implemented in IsaacSim API).

Code for RPL - per-terrain experts RL training, multi-depth distillation, and sim2sim, built on Holosoma.

Capabilities:

• Bidirectional Multi-Terrain Locomotion - walks forward and backward across stairs, stepping stones, and slopes.
• Whole-Body Teleoperation - tracks commanded base velocity and end-effector pose, with base height and torso orientation adjustable during stance.
• Force-Adaptive Payload Carrying - compensates for unknown external forces at the end effectors.

Repo layout

Path	What it holds
thirdparty/holosoma/	Holosoma core (git submodule): warp ray-cast multi-depth rendering, PPO-MA, DAgger-MA, ViT / CNN modules, terrains
rpl_training/	Everything outside Holosoma core: loco-manipulation env, manager terms, presets, launchers, tests
rpl_training/data/	ACCAD upper-body motion clips and the four teachers (models/)
rpl_sim2sim/	Sim2sim: MuJoCo environment (sim side), the depth student for holosoma_inference (policy side), reference student exports

Setup

Linux with NVIDIA GPUs.

git clone --recurse-submodules https://github.com/amazon-far/RPL.git
cd RPL

The sim2sim demo and training each install their own conda envs.

Sim2sim (MuJoCo)

rpl_sim2sim.mp4

Runs the depth student in the MuJoCo sim2sim environment in rpl_sim2sim/ (video). The sim and the policy run as two processes communicating over Unitree sdk2py (DDS), with depth images passed through shared memory.

1. Install

bash rpl_sim2sim/setup_env.sh

This builds two conda envs under ~/.holosoma_deps/miniconda3 (override with CONDA_ROOT):

Env	Role	Override
rplmujoco	Sim: MuJoCo, depth renderer, Xbox pad, Unitree SDK bridge	SIM_ENV
rplinference	Policy: holosoma_inference + the RPL depth student	POLICY_ENV

The script needs an existing conda at CONDA_ROOT. If you have none, run thirdparty/holosoma/scripts/setup_mujoco.sh once first.

2. Launch

Tip

We highly recommend an Xbox controller for sim2sim. Plug it into the host running run_sim.sh. The keyboard works, but it's harder to steer smoothly.

Start the sim first, then the policy, each in its own terminal:

# Terminal 1: sim (terrain course, Xbox pad on this host)
bash rpl_sim2sim/run_sim.sh

# Terminal 2: policy (shipped student)
bash rpl_sim2sim/run_policy.sh

Goal	Command
Run your own student	run_policy.sh --task.model-path /path/model_XXXXX.onnx
Use the keyboard instead of an Xbox pad	run_sim.sh --no-joystick (see Controls)

run_sim.sh --help and run_policy.sh --help list the rest.

3. Walk

1. Release the elastic band: press 9 in the MuJoCo window.
2. Start the policy right away: start on the pad (] on the keyboard).
3. Switch to walking: R2 (=).
4. Drive with the sticks (w/s/a/d, q/e).

Controls

Policy, Xbox pad:

Input Action start Start the policy A + X Init pose R2 Toggle stand / walk Left stick Forward speed Right stick Yaw L2 Zero velocity B + Y Stop B + d-pad up / down Base height Y / select + d-pad Waist R1 + d-pad Hands forward / back, apart / together X + d-pad up / down Hands up / down

Policy, keyboard (instead of the Xbox pad)

Run run_sim.sh --no-joystick and run_policy.sh --task.velocity-input keyboard --task.state-input keyboard:

Key	Action
]	Start the policy
i	Init pose
=	Toggle stand / walk
w / s / a / d	Linear velocity
q / e	Angular yaw velocity
o	Stop

MuJoCo window:

Key	Action
9	Release the elastic band
7 / 8	Band length
3 - 6	Move the band anchor
R	Toggle video recording (logs/sim2sim/simulation.mp4)

Training

Install

bash rpl_training/training_runs/setup_env.sh

This builds the IsaacGym conda env hsgym. Optional environment variables:

Variable	Effect
ENV_NAME	Conda env name (default hsgym)
HOLOSOMA_LOCAL_DIR	Use another Holosoma checkout instead of thirdparty/holosoma

The launchers log to W&B by default. Set LOGGER=disabled to keep logs local under logs/, or WANDB_PROJECT / WANDB_ENTITY / RUN_NAME to pick the W&B destination. NGPUS sets the number of GPUs, and extra arguments pass through to rpl_training.train_agent.

1. Train the RL experts

We recommend training the experts on 4 GPUs.

NGPUS=4 EXPERT=stones bash rpl_training/training_runs/run_expert.sh --training.name stones-expert
NGPUS=4 EXPERT=stairs bash rpl_training/training_runs/run_expert.sh --training.name stairs-expert
NGPUS=4 EXPERT=slopes bash rpl_training/training_runs/run_expert.sh --training.name slopes-expert

2. Distill the multi-depth student

We recommend training the student on 8 GPUs.

NGPUS=8 bash rpl_training/training_runs/run_distill.sh --training.name rpl-student

It loads the shipped teachers by default. To distill your own experts, pass either:

# A directory with the same per-terrain layout as rpl_training/data/models
--algo.config.ckpt-teacher DIR

# Or one W&B checkpoint per terrain (slopes, stairs_up, stairs_down, stepping_stones_columns)
--algo.config.ckpt-teacher-wandb.slopes wandb://ENTITY/PROJECT/RUN_ID/model_XXXXX.pt

Distillation runs export model_XXXXX.onnx next to each checkpoint for sim2sim.

Citation

@article{zhang2026rpl,
  title   = {RPL: Learning Robust Humanoid Perceptive Locomotion on Challenging Terrains},
  author  = {Zhang, Yuanhang and Seo, Younggyo and Chen, Juyue and Yuan, Yifu and Sreenath, Koushil and Abbeel, Pieter and Sferrazza, Carmelo and Liu, Karen and Duan, Rocky and Shi, Guanya},
  journal = {arXiv preprint arXiv:2602.03002},
  year    = {2026}
}

Related: FALCON

License

Apache-2.0, see LICENSE.