feat: G1 per-joint kp/kd domain randomization (issue #129) by lemon-star608 · Pull Request #259 · Motphys/UniLab

lemon-star608 · 2026-04-16T02:34:17Z

Summary

为 G1 joystick（PPO / SAC）引入 per-joint kp/kd 域随机化，与 Go2 架构对齐，解决 issue Refactor: G1 joystick 统一采用 MuJoCo kp/kd 域随机化方案，移除 backend_overrides 临时方案 #129
在 SimBackend 上新增 get_actuator_gains() 作为正式可选能力合约（NotImplementedError 默认值，与 get_physics_state() 同一模式），MuJoCoBackend 实现该方法从模型状态读取 per-joint kp/kd 数组
在 LocomotionDRProvider 上新增 _get_base_actuator_gains() 钩子方法（与 _get_qvel_limit / _build_extra_info_updates 同一模式），默认返回 (None, None) 回退到标量 ControlConfig.Kp/Kd，Go1/Go2 行为不变
G1JoystickDomainRandomizationProvider 重写该钩子，在 env init 冷路径通过正式合约 backend.get_actuator_gains() 读取并缓存 per-joint 基准 gains；DR 关闭时不调用 get_actuator_gains()，Motrix backend 若被误配置为开启则在初始化阶段显式抛出 NotImplementedError
dr_utils.build_common_reset_randomization 和 validate_common_reset_randomization 新增 base_kp/base_kd 关键字参数，移除原有 getattr(env, "_backend") / hasattr(backend, "get_actuator_gains") 隐式 backend 探测代码；新参数均为可选默认 None，所有现有调用方完全向后兼容
G1 Motrix 配置通过 randomize_kp: false / randomize_kd: false 显式禁用 kp/kd DR

Linked Work

Issue: Refactor: G1 joystick 统一采用 MuJoCo kp/kd 域随机化方案，移除 backend_overrides 临时方案 #129

Validation

make check
uv run pytest -m "not slow"
MuJoCo backend：kp/kd DR 开启，per-joint gains 正确缓存并随机化（29 关节，范围 [14.3, 99.1]）
Motrix backend：randomize_kp/kd: false，provider 无缓存 gains，payload 不含 kp/kd
Motrix backend 误配置为开启：初始化阶段显式抛出 NotImplementedError

Commands actually run:

uv run pyright src/unilab/base/backend/base.py \
               src/unilab/base/backend/mujoco_backend.py \
               src/unilab/dr/dr_utils.py \
               src/unilab/envs/locomotion/common/dr_provider.py \
               src/unilab/envs/locomotion/g1/joystick.py \
               src/unilab/envs/locomotion/g1/joystick_sac.py
uv run pytest -m "not slow"

Impact

Backend impact: both（MuJoCo 新增 get_actuator_gains() 实现；Motrix 继承 NotImplementedError 默认值，行为不变）
Platform impact: both
Training effect expected: yes（MuJoCo backend 下 kp/kd DR 生效；Motrix 行为不变）

Checklist

Added or updated tests where needed
Updated docs if behavior or workflow changed
Linked the driving issue
Noted any follow-up work explicitly

Copilot

Pull request overview

This PR aligns G1 joystick (PPO/SAC) with the Go2-style kp/kd domain randomization flow by introducing per-joint actuator gain randomization via a formal backend capability (get_actuator_gains()), and wiring that through the locomotion DR provider/reset payload generation.

Changes:

Add an optional SimBackend.get_actuator_gains() contract and implement it for MuJoCoBackend to read per-actuator kp/kd from the model.
Introduce G1DomainRandConfig + G1 DR provider gain caching, and pass cached base kp/kd into common reset randomization utilities.
Update Motrix task configs to explicitly disable kp/kd randomization (avoiding unsupported-backend initialization errors).

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/unilab/envs/locomotion/g1/joystick.py	Adds G1-specific DR config/provider hook; conditionally caches per-joint gains from backend during env init.
src/unilab/envs/locomotion/g1/joystick_sac.py	Switches SAC config to G1 DR config and mirrors the gain-caching DR initialization path.
src/unilab/envs/locomotion/common/dr_provider.py	Adds `_get_base_actuator_gains()` hook and threads base kp/kd into validate/reset-plan generation.
src/unilab/dr/dr_utils.py	Extends common reset randomization builder/validator with optional `base_kp/base_kd` and removes implicit backend probing.
src/unilab/base/backend/base.py	Defines optional backend API `get_actuator_gains()` with a default `NotImplementedError`.
src/unilab/base/backend/mujoco_backend.py	Implements `get_actuator_gains()` by reading MuJoCo actuator gain/bias parameters.
conf/ppo/task/g1_joystick/motrix.yaml	Disables kp/kd DR under Motrix for PPO.
conf/offpolicy/task/sac/g1_sac/motrix.yaml	Disables kp/kd DR under Motrix for SAC.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-16T03:50:21Z

+        if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:
+            base_kp, base_kd = backend.get_actuator_gains()
+            dr_provider = G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)
+        else:
+            dr_provider = G1JoystickDomainRandomizationProvider()
+        self._init_domain_randomization(dr_provider)



The domain-randomization provider selection logic is duplicated here and in g1/joystick_sac.py. To reduce the chance of future drift (e.g., adding a new DR term in one env but not the other), consider factoring this into a shared helper (or a small method on the base env) that returns the correctly-configured provider.

Suggested change

if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:

base_kp, base_kd = backend.get_actuator_gains()

dr_provider = G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)

else:

dr_provider = G1JoystickDomainRandomizationProvider()

self._init_domain_randomization(dr_provider)

dr_provider = self._create_domain_randomization_provider(backend)

self._init_domain_randomization(dr_provider)

def _create_domain_randomization_provider(

self, backend: Any

) -> G1JoystickDomainRandomizationProvider:

if self._cfg.domain_rand.randomize_kp or self._cfg.domain_rand.randomize_kd:

base_kp, base_kd = backend.get_actuator_gains()

return G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)

return G1JoystickDomainRandomizationProvider()

Copilot · 2026-04-16T03:50:21Z

        self._init_reward_functions()
-        self._init_domain_randomization(G1JoystickDomainRandomizationProvider())
+        if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:
+            base_kp, base_kd = backend.get_actuator_gains()


For clarity and to avoid relying on a local variable after initialization, consider calling self._backend.get_actuator_gains() instead of backend.get_actuator_gains() here. It makes it explicit that the gains are being read from the env’s active backend instance.

Suggested change

base_kp, base_kd = backend.get_actuator_gains()

base_kp, base_kd = self._backend.get_actuator_gains()

Copilot · 2026-04-16T03:50:21Z

+        if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:
+            base_kp, base_kd = backend.get_actuator_gains()
+            dr_provider = G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)
+        else:
+            dr_provider = G1JoystickDomainRandomizationProvider()
+        self._init_domain_randomization(dr_provider)



The domain-randomization provider selection logic is duplicated with g1/joystick.py. Consider extracting a shared helper so PPO/SAC variants stay in sync when DR behavior changes (e.g., additional reset terms or different gain-caching rules).

Suggested change

if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:

base_kp, base_kd = backend.get_actuator_gains()

dr_provider = G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)

else:

dr_provider = G1JoystickDomainRandomizationProvider()

self._init_domain_randomization(dr_provider)

dr_provider = self._create_domain_randomization_provider(cfg, backend)

self._init_domain_randomization(dr_provider)

def _create_domain_randomization_provider(self, cfg: G1JoystickSACCfg, backend):

"""Create the DR provider while staying aligned with PPO/shared logic when available."""

parent_factory = getattr(super(), "_create_domain_randomization_provider", None)

if callable(parent_factory):

return parent_factory(cfg, backend)

if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:

base_kp, base_kd = backend.get_actuator_gains()

return G1JoystickDomainRandomizationProvider(base_kp=base_kp, base_kd=base_kd)

return G1JoystickDomainRandomizationProvider()

Copilot · 2026-04-16T03:50:22Z

        self._init_reward_functions()
-        self._init_domain_randomization(G1JoystickDomainRandomizationProvider())
+        if cfg.domain_rand.randomize_kp or cfg.domain_rand.randomize_kd:
+            base_kp, base_kd = backend.get_actuator_gains()


For readability, prefer self._backend.get_actuator_gains() over the local backend variable here, so it’s clear the gains are read from the env’s configured backend instance.

Suggested change

base_kp, base_kd = backend.get_actuator_gains()

base_kp, base_kd = self._backend.get_actuator_gains()

TATP-233 · 2026-04-16T06:31:32Z

再补充一下修改前后的训练对比曲线

lemon-star608 · 2026-04-16T06:42:10Z

紫色为仅开启kdkpDR后的曲线

TATP-233 · 2026-04-16T10:43:48Z

mujoco_uni.batch_env
pool.get_field(env_id, name) -> np.ndarray
提供了批量获取DR参数属性的功能，替换成这个接口

feat: G1 per-joint kp/kd domain randomization (issue #129)

Closes Motphys#1617. UniSim 1.7.x accepts a bounded, readback-validated PhysX solver override set on the IsaacSim backend (unilabsim/unisim#251, Motphys#259), but the standard UniLab path (registry.make -> make_manager_based_rl_env -> create_backend) had no channel for those knobs: task owners could not declare the solver table and the backend silently ran PhysX defaults. - EnvCfg gains the six solver fields, following the per-backend knob convention (Optional, None keeps the PhysX scene defaults): isaacsim_solver_position_iteration_count (int >= 1), isaacsim_solver_velocity_iteration_count (int >= 0), isaacsim_bounce_threshold_velocity / isaacsim_contact_offset / isaacsim_rest_offset / isaacsim_max_depenetration_velocity (non-negative floats). EnvCfg.validate mirrors the PhysxSolverConfig bounds so owner configs fail fast. - env_backend_kwargs forwards all six to the UniSim factory under the same names. - Tests: field->kwargs mapping assertions (values forwarded verbatim, defaults stay None) and validation rejection coverage. - Changelog entry (shared page, Added).

feat: G1 per-joint kp/kd domain randomization (issue #129)

2fde3a6

lemon-star608 force-pushed the g1_deploy branch from 8c338f5 to 2fde3a6 Compare April 16, 2026 02:37

lemon-star608 requested a review from Copilot April 16, 2026 03:44

Copilot started reviewing on behalf of lemon-star608 April 16, 2026 03:44 View session

Copilot AI reviewed Apr 16, 2026

View reviewed changes

TATP-233 merged commit df6942f into main Apr 16, 2026
9 checks passed

TATP-233 deleted the g1_deploy branch April 16, 2026 06:32

TATP-233 reviewed Apr 16, 2026

View reviewed changes

TATP-233 added a commit that referenced this pull request May 4, 2026

Merge pull request #259 from unilabsim/g1_deploy

12727ff

feat: G1 per-joint kp/kd domain randomization (issue #129)

TATP-233 added a commit that referenced this pull request May 30, 2026

Merge pull request #259 from unilabsim/g1_deploy

c130d82

feat: G1 per-joint kp/kd domain randomization (issue #129)

TATP-233 added a commit that referenced this pull request Sep 4, 2026

Merge pull request #259 from unilabsim/g1_deploy

fa59c9d

feat: G1 per-joint kp/kd domain randomization (issue #129)

This was referenced Sep 21, 2026

feat(isaacsim)：通过 EnvCfg 后端接线暴露 UniSim 的 PhysX solver 旋钮 #1617

Closed

feat(isaacsim)：通过 EnvCfg 后端接线暴露 UniSim 的 PhysX solver 旋钮 #1620

Merged

Navigation Menu

Repository navigation

feat: G1 per-joint kp/kd domain randomization (issue #129) - #259

feat: G1 per-joint kp/kd domain randomization (issue #129)#259
TATP-233 merged 1 commit into
mainfrom
g1_deploy

lemon-star608 commented Apr 16, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

Copilot AI Apr 16, 2026

Uh oh!

TATP-233 commented Apr 16, 2026

Uh oh!

Uh oh!

lemon-star608 commented Apr 16, 2026 •

edited

Loading

Uh oh!

TATP-233 Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	base_kp, base_kd = backend.get_actuator_gains()
	base_kp, base_kd = self._backend.get_actuator_gains()

Conversation

lemon-star608 commented Apr 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Linked Work

Validation

Impact

Checklist

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

TATP-233 commented Apr 16, 2026

Uh oh!

Uh oh!

lemon-star608 commented Apr 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

TATP-233 Apr 16, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

lemon-star608 commented Apr 16, 2026 •

edited

Loading

lemon-star608 commented Apr 16, 2026 •

edited

Loading