Skip to content
JihoonPark20Public

About

Rhythmic Information States: Separating Historical Timing Priors from Proprioceptive History for Quadrupedal Locomotion

Resources

Stars

2 stars

Watchers

0 watching

Forks

Latest commit

 

History

3 Commits

Folders and files

Repository files navigation

Rhythmic Information States

Separating Historical Timing Priors from Proprioceptive History for Quadrupedal Locomotion

Jihoon Park

Paper · Project Page · Code

Deployed RIS policy on Unitree Go2 in MuJoCo, switching Trot, Pace, Bound and Pronk

The deployed Full RIS policy, unmodified, running sim-to-sim in MuJoCo at 1× speed. The gait command switches Trot → Pace → Bound → Pronk at 1.0 m/s and 3 Hz. Bottom: desired stance (grey) and measured foot contact (black).

Overview

Clock-conditioned locomotion policies often feed past clock signals into the same history window as proprioception. Our attribution analysis shows these historical clocks become the dominant input pathway, and the policy loses commanded gait realization when they are removed.

RIS keeps the current clock, command and proprioception explicit. It encodes only proprioceptive history into a compact stochastic latent, trained with value learning, next-observation prediction and a variational bottleneck. At deployment only the actor and the history encoder remain.

 o_{t-19..t} ──► history encoder ──► z_t ─┐
                                          ├──► actor ──► a_t
 [ o_t , u_t , c_t ] ─────────────────────┘
   proprio  command  current clock         (no past clocks in the encoder)

Results

Narrow-domain gait controllability, 3 seeds × 128 episodes per gait.

Method GQ ↑ (ID) GQ ↑ (held-out) JS % ↑ (ID) JS % ↑ (held-out)
WTW 0.62 0.58 46.6 26.4
WTW − clock history 0.19 0.25 14.3 12.8
RIS + clock history 0.87 0.78 88.3 68.5
Full RIS 0.87 0.77 87.7 71.4

Removing historical clocks costs WTW −0.44 GQ but costs RIS nothing. RIS also expands the gait-conditioned command curriculum earlier for the same number of environment steps.

Installation

Requires Ubuntu, an NVIDIA GPU, Python 3.8 and Isaac Gym Preview 4.

conda create -n ris python=3.8 && conda activate ris
pip install -e /path/to/isaacgym/python   # also installs PyTorch
git clone https://github.com/JihoonPark20/RIS.git && cd RIS
pip install -e rsl_rl && pip install -e .

Training

python legged_gym/scripts/train.py --task go2 --headless --experiment_name ris

--num_envs, --seed, --max_iterations and --resume override the config. --no_wandb turns off logging.

Environment

Robot Unitree Go2, 12 DoF, joint-position residual control
Control 50 Hz policy (sim 200 Hz, decimation 4), PD Kp 25 / Kd 0.6
Policy input proprioception 42 · command 12 · per-foot sin/cos clock 8
History encoder 20 steps of proprioception only, stochastic latent, β = 1e-4
Privileged targets base linear velocity, foot contacts (training only)
Gaits Trot, Pace, Bound, Pronk; 2–4 Hz; adaptive command curriculum
Domain randomization friction, restitution, payload, gravity, motor strength/offset, Kp/Kd

All of these are set in legged_gym/envs/go2/go2_config.py.

legged_gym/envs/go2/        Go2 task and reward config
legged_gym/envs/base/       environment, gait clock, curriculum
rsl_rl/rsl_rl/modules/      actor-critic and RIS history encoder
rsl_rl/rsl_rl/algorithms/   PPO with the encoder objective

Acknowledgements

Built on legged_gym, rsl_rl and walk-these-ways.

Citation

@misc{park2026ris,
  title  = {Rhythmic Information States: Separating Historical Timing Priors
            from Proprioceptive History for Quadrupedal Locomotion},
  author = {Park, Jihoon},
  year   = {2026}
}

About

Rhythmic Information States: Separating Historical Timing Priors from Proprioceptive History for Quadrupedal Locomotion

Resources

Stars

2 stars

Watchers

0 watching

Forks

Contributors

Languages