Changelog Highlights

This page summarizes the most important recent changes. For the full history, see CHANGELOG.md in the repository root.

v0.7.0 · 2026-04-24

PPO training divergence fixes

Five root causes of training instability were identified and resolved:

  • init_std 1.0 → 0.1 — old value caused ±0.08 rad per-step noise, short episodes, and std growing to 12+ during training
  • num_steps_per_env 24 → 2048 — 24 steps (0.48s) was shorter than one gait cycle (0.72s); GAE could not see a complete stride
  • max_iterations 500 → 30000 (flat), 1500 → 5000 (rough) — biped locomotion needs far more iterations than the template default
  • clip_actions None → 10.0 — unbounded outputs with large std immediately toppled the robot
  • num_mini_batches 4 → 32 — matches the larger rollout buffer
  • reset_robot_joints position_range (0.85, 1.15) → (1.0, 1.0) — scale-based perturbation is meaningless for near-zero joints
  • RayCasterCfg: attach_yaw_only=True → ray_alignment="yaw" (API update)

v0.6.0 · 2026-04-24

Housekeeping and docs refresh

This release cleaned up leftover template code and updated the documentation:

  • fixed list_envs.py table display — class entry points now render as module.ClassName with capped column widths,
  • removed the unused anymal_d task config (template scaffold, never used),
  • bumped Isaac Sim requirement to 5.1.0 and Isaac Lab to Latest,
  • added Feishu internal setup guide link to getting-started,
  • added a new Tested Environments page documenting the validated hardware and software stack.

v0.5.0 · 2026-04-19

Gait quality improvements

This release focused on two visible issues:

  • the robot taking tiny, unclear steps,
  • insufficient knee lift during swing.

Key updates included:

  • larger hip swing amplitude,
  • larger knee swing amplitude,
  • stronger swing-phase knee scaling,
  • shorter stance ratio,
  • larger residual action budget.

It also increased reward targets for hip alternation and knee flexion, while making air-time rewards easier to trigger.

v0.4.0 · 2026-04-18

Stability and symmetry fixes

This release fixed:

  • training crashes caused by invalid standard deviation handling,
  • asymmetric limping behavior,
  • weak alternation incentives.

Key updates included:

  • migration to distribution_cfg with std_type="log",
  • stronger hip_alternation and knee_flexion rewards,
  • a new knee symmetry penalty,
  • stronger feet air-time encouragement.

v0.3.0 · 2026-04-18

play.py compatibility and anti-circling reward updates

This release addressed:

  • rsl_rl >= 5.0 compatibility in play.py,
  • export path changes for TorchScript and ONNX,
  • observation API updates,
  • circling behavior caused by missing yaw-rate punishment.

It also:

  • lowered the curriculum starting speed,
  • increased the max curriculum speed,
  • increased training iterations,
  • added a small entropy bonus.

v0.2.0 · 2026-04-16

Core Guguji locomotion stack landed

Added:

  • flat and rough environment configs,
  • PPO configs,
  • Gym registrations,
  • reference gait action,
  • gait phase observations,
  • velocity command curriculum.

v0.1.0 · 2026-04-16

Repository bootstrap

Initial repository setup included:

  • Isaac Lab extension initialization,
  • English and Chinese readmes,
  • migration context documentation,
  • root changelog.