[Proposal] Optimization of default built solver_cfg for dexforceW1

[yhnsu]

Proposal: Refactor DexforceW1Cfg solver initialization (PytorchSolver vs SRSSolver overwrite)

File: embodichain/lab/sim/robots/dexforce_w1/cfg.py
Class: DexforceW1Cfg

---

Current behavior

@configclass
class DexforceW1Cfg(RobotCfg):
    ...

    @classmethod
    def from_dict(
        cls, init_dict: Dict[str, str | float | tuple | dict]
    ) -> DexforceW1Cfg:
        ...
        init_dict_m = init_dict.copy()
        version = init_dict_m.get("version", "v021")
        arm_kind = init_dict_m.get("arm_kind", "anthropomorphic")
        with_default_eef = init_dict_m.get("with_default_eef", True)
        init_dict_m.pop("version", None)
        init_dict_m.pop("arm_kind", None)
        init_dict_m.pop("with_default_eef", None)

        # 1) Build base cfg via utils (this already initializes PytorchSolver solver_cfg based on URDF + control_parts)
        cfg: DexforceW1Cfg = cls()._build_default_cfg(
            version=version, arm_kind=arm_kind, with_default_eef=with_default_eef
        )

        # 2) Merge default physics configs
        default_physics_cfgs = cls()._build_default_physics_cfgs(
            arm_kind=arm_kind, with_default_eef=with_default_eef
        )
        for key, value in default_physics_cfgs.items():
            setattr(cfg, key, value)

        # 3) Build SRSSolver configs for both arms
        default_solver_cfg = cls()._build_default_solver_cfg(
            is_industrial=(arm_kind == "industrial")
        )

        # 4) ⚠️ This line overwrites the PytorchSolver-based solver_cfg created in _build_default_cfg
        cfg.solver_cfg = default_solver_cfg

        # 5) Finally merge external overrides
        cfg = merge_robot_cfg(cfg, init_dict_m)

        return cfg

Notes:

• _build_default_cfg() internally calls build_dexforce_w1_cfg() (in dexforce_w1/utils.py), which already builds a complete solver_cfg based on the assembled URDF and control_parts, including:

  • full_body
  • left_arm_to_torso
  • right_arm_to_torso
  • other whole-body / torso-related control parts as PytorchSolverCfg.

• _build_default_solver_cfg() only returns SRSSolver configs for the two 7-DoF arms:

@staticmethod
def _build_default_solver_cfg(is_industrial: bool) -> SolverCfg:
    from embodichain.lab.sim.solvers import SRSSolverCfg
    from embodichain.lab.sim.robots.dexforce_w1.params import W1ArmKineParams

    if is_industrial:
        ...
    else:
        ...

    return {
        "right_arm": SRSSolverCfg(...),
        "left_arm": SRSSolverCfg(...),
    }

As a result:

• build_dexforce_w1_cfg() first initializes a rich set of PytorchSolver entries (full_body / *_arm_to_torso / ...).
• Then cfg.solver_cfg = default_solver_cfg replaces the entire dict with a much smaller one that only contains "left_arm" and "right_arm" SRSSolver entries.
• All PytorchSolver-based configs built in utils.build_dexforce_w1_solver_cfg are effectively discarded; their initialization work becomes meaningless and the final behavior is non‑obvious to users.

---

Why this is problematic

DexforceW1 is designed to support two complementary IK solver families:

• URDF-based PytorchSolver:

  • Used for whole-body / torso-related control parts such as full_body, left_arm_to_torso, right_arm_to_torso, etc.

• DH-based SRSSolver:

  • Used for the 7-DoF arms: left_arm, right_arm (high-quality, analytic-like IK).

The current from_dict() implementation:

• discards the PytorchSolver-based entries created in _build_default_cfg / build_dexforce_w1_cfg,
• keeps only the SRSSolver entries for left_arm / right_arm,
• breaks the intended layering: “whole-body PytorchSolver” + “per-arm SRSSolver”.

It also makes the code path harder to reason about:
logging solver_cfg right after build_dexforce_w1_cfg() shows rich PytorchSolver configs, but the final DexforceW1Cfg.solver_cfg that Robot.init_solver sees is a different, reduced set.

---

Proposed improvement

Goals:

• Preserve the PytorchSolver-based solver configs produced by build_dexforce_w1_cfg() for:

  • full_body
  • left_arm_to_torso
  • right_arm_to_torso
  • etc.

• Still provide SRSSolver-based defaults for:

  • "left_arm"
  • "right_arm"

• Allow external robot.solver_cfg in user configs to override individual fields (e.g., tcp, ik_nearest_weight) instead of replacing the entire solver_cfg dict.

Suggested change (only in the solver initialization part of from_dict):

@classmethod
def from_dict(
    cls, init_dict: Dict[str, str | float | tuple | dict]
) -> DexforceW1Cfg:
    ...
    cfg: DexforceW1Cfg = cls()._build_default_cfg(
        version=version, arm_kind=arm_kind, with_default_eef=with_default_eef
    )

    default_physics_cfgs = cls()._build_default_physics_cfgs(
        arm_kind=arm_kind, with_default_eef=with_default_eef
    )
    for key, value in default_physics_cfgs.items():
        setattr(cfg, key, value)

    # Build SRSSolver defaults for left/right arms
    default_solver_cfg = cls()._build_default_solver_cfg(
        is_industrial=(arm_kind == "industrial")
    )

    # ✅ Proposed: merge instead of overwrite.
    # If build_dexforce_w1_cfg did not set solver_cfg, just use SRSSolver defaults.
    # Otherwise, only update "left_arm"/"right_arm" with SRSSolver, keeping
    # full_body / left_arm_to_torso / right_arm_to_torso PytorchSolver entries.
    if cfg.solver_cfg is None:
        cfg.solver_cfg = default_solver_cfg
    else:
        cfg.solver_cfg.update(default_solver_cfg)

    cfg = merge_robot_cfg(cfg, init_dict_m)
    return cfg

With this:

• build_dexforce_w1_cfg() + build_dexforce_w1_solver_cfg() are responsible for:

  • default control_parts layout (torso, head, left_arm, right_arm, full_body, ...);
  • URDF-based PytorchSolverCfg for full_body, left_arm_to_torso, right_arm_to_torso, etc.

• _build_default_solver_cfg() is responsible only for:

  • providing SRSSolverCfg for "left_arm" and "right_arm" (used to overwrite or supplement those keys).

• External init_dict["solver_cfg"]:

  • can safely override fields (e.g., tcp, ik_nearest_weight) via merge_robot_cfg without unintentionally wiping out all other solver entries.

---

Summary

Currently, the PytorchSolver-based solver_cfg built in _build_default_cfg (via build_dexforce_w1_cfg) is entirely overwritten in from_dict by _build_default_solver_cfg, making the earlier initialization effectively useless.

By changing the logic to merge (cfg.solver_cfg.update(default_solver_cfg)) instead of blindly overwriting, we can:

• keep whole-body / torso PytorchSolver IK intact,
• still use SRSSolver for arm-specific IK,
• and make the configuration behavior much easier to understand and extend.