PPO policy checkpoints for the SDC (serpenoid curve control) snake-robot research line, trained in IsaacLab with rsl_rl. This repository is the single archive for the whole line: one directory per training round, each holding the final checkpoints together with the exact env.yaml / agent.yaml they were trained with.

Originally published as snake-rl-sdc3d20-20260727; renamed into this repository on 2026-07-27 (the old URL redirects here).

Two PPO policies trained in IsaacLab on a 20-DoF snake robot with a 12-dimensional 3D SDC action space (yaw + pitch amplitude groups, shared Gaussian window, inter-plane phase offset). Both were trained for 3000 iterations at 4096 environments over a single 160-tile terrain grid covering 8 families.

Code commit: cb903bf on branch feature/sdc3d-action-space.

sdc3d20_20260727/attoff/model_2999.pt     attitude penalties disabled  (roll_rate = 0.0,  vz = 0.0)
sdc3d20_20260727/attoff/env.yaml          environment config as run
sdc3d20_20260727/attoff/agent.yaml        PPO config as run
sdc3d20_20260727/atton/model_2999.pt      attitude penalties enabled   (roll_rate = -1.0, vz = -3.0)
sdc3d20_20260727/atton/env.yaml
sdc3d20_20260727/atton/agent.yaml

The two runs are identical except for those two scalars: same seed (42), same task Uniteboa-SDC3D20-AllTerrains-v0, same 4096 environments, same 3000 iterations, progress_cap_to_cmd = true in both.

12 continuous dimensions per step:

Joint targets follow a travelling serpenoid wave. Even joints are yaw, odd joints are pitch. The pitch plane reuses the same Gaussian amplitude map as yaw, so the existing sdc_basis ablation applies identically to both planes. phi = +/- pi/2 corresponds to sidewinding, phi = 0 to sinus lifting.

4096 environments, 2000 steps, measurement window [200, first reset), net progress projected onto the spawn heading. The 2D column is the previous 8-dimensional SDC champion, measured under the same protocol at command -0.035 m/s.

Fleet mean net progress over 36 s at command -0.035 m/s:

Per family, restricted to the families where the comparison was verified directly:

Fleet speed against commanded speed:

Neither policy is a clean win.

att-ON collapsed. The pitch amplitude weight sits at -1.56 with 100 percent of environments at the lower bound, and the yaw amplitude collapsed as well, to -1.68. The policy froze rather than reverting to a 2D gait; 68.6 percent of environments on stairs are completely stationary, and fleet progress falls below the 2D baseline. Penalising vertical velocity removes the very mechanism 3D locomotion depends on. Roll rate is the tumbling-exploit channel and should be penalised separately from vz.

att-OFF is command-blind and partly gravity-assisted. Speed is essentially constant across three different commanded speeds, so the policy does not track the command at all. On stairs it descends 2.638 m in 36 s and on slopes 0.844 m, so a substantial part of the headline progress is falling rather than locomotion.

The defensible 3D gain is narrow but real: on flat ground, where no descent is available, att-OFF reaches 3.193 m against 0.756 m for 2D at dz = -0.001. Pole forest regresses to 0.406 m against 2.657 m, so lifting destroys the pole-contact propulsion the 2D gait relies on.

Both policies keep |phi| > 0.7 at 0.0 percent of environments, att-OFF averaging 0.44 rad, so the gain does not come from sidewinding but from a mild yaw/pitch phase offset, morphologically closer to sinus lifting combined with lateral undulation.

Efficiency is worse in 3D: att-OFF spends 30.5 m of body path for 3.38 m of net displacement (11 percent), against 6.29 m for 1.078 m (17 percent) in 2D.

Single seed (42) per arm. No seed replication was run.

IsaacLab / Isaac Sim 5.1, direct RL environment, rsl_rl PPO. Isotropic friction (static = dynamic = 0.2, combine mode min), no passive wheels, scales or anisotropy. Control step 0.02 s (decimation 4, sim dt 1/200), episode 2000 steps or 40 s.

Adaptive Learning System Control Laboratory, Okayama University.