Background
PR #699 (fixing #638: payload(0) wasn't a no-op) landed a real fix, but review surfaced three related design questions it deliberately left out of scope. Grouped here since they're all facets of the same underlying question: what does payload() actually own, and does that model hold up once robots stop being simple serial chains?
Confirmed: payload doesn't survive copy()
Robot.payload() now remembers a link's pre-payload m/r/I in a private link._payload_own attribute so payload(0) can restore it. That attribute is Robot-level bookkeeping, but it's stored on the Link object -- and Link.__deepcopy__ (correctly, from Link's own point of view) only reconstructs the named dynamic parameters it knows about, so _payload_own never survives a copy. Reproduced directly:
p560 = rp.models.DH.Puma560()
p560.payload(2.5, [0.02, -0.03, 0.1])
p560_copy = p560.copy()
p560_copy.payload(0) # supposed to remove the payload -- silently does nothing
p560_copy.rne(...) after payload(0) is bit-for-bit identical to with the payload, not the bare-link baseline. No exception, no warning -- the copy's payload is permanently baked into what it now believes is the link's own mass. PR #699 documents this as a known limitation in its docstring, but "known" isn't "fixed," and the failure mode is silent wrong dynamics, not an error.
This is the same root pattern as #577 (Link.__deepcopy__ silently drops coal collision geometry): Link.__deepcopy__ only carries a fixed set of named parameters, so anything stored as an ad-hoc attribute on Link -- regardless of which class conceptually owns it -- disappears across a copy with no signal to the caller.
Proper fix: move the "a payload is attached, here's what to subtract" bookkeeping onto Robot/BaseRobot itself (e.g. self._payload: tuple[Link, float, NDArray, NDArray] | None), the same way Robot.__deepcopy__/DHRobot.__deepcopy__ already explicitly re-attach other instance-level state beyond the constructor kwargs (configs, ikine_a). That keeps Link as a dumb parameter container and fixes the copy bug at the source, rather than teaching Link.__deepcopy__ about a Robot-level concept smuggled onto it.
payload() has no answer for multiple end-effectors
For a robot with len(ee_links) > 1, #699 falls back to the pre-existing behaviour of using links[n - 1] -- whichever link happens to land last in the flattened links list. For a genuine tree that's arbitrary: it has no necessary physical relationship to either branch, and is fragile to construction/traversal order.
The idiomatic fix is the end= parameter pattern already used throughout the kinematics API (fkine, jacob0, etc., via self._getlink(end, default)): give payload() an end= parameter, defaulting to ee_links[0] only when there's exactly one, and require it explicitly (raise, don't guess) when there's more than one.
Robot.rne()'s branching support has no test coverage
While investigating the above, confirmed empirically that Robot.rne() does correctly handle a genuine kinematic tree (multiple independent moving-joint chains sharing a parent, not just static leaf links) -- built a synthetic two-armed robot and verified gravity-torque superposition holds and there's no cross-branch leakage. The group_of_link_idx/backward-force-summation code in Robot.rne() clearly was written deliberately (extensive comments referencing #636/#483/#571), but there is currently no test in the repo that exercises it -- the existing "branched robot" tests (test_dotfile2, test_graph_multiple_end_effectors, YuMi's fkine/jacob0 tests) only touch kinematics/graph rendering, and #699's own test_payload_several_end_effectors has two static tip links off one joint (branching in link topology, not in DOF), so it never reaches the force-summation-at-shared-parent path.
Worth a regression test using something like the synthetic two-armed robot above (or YuMi directly), independent of the payload/end= work.
Scope
Not blocking #699, which fixes a real, independently-valuable bug (#638) on its own. Related: #577 (same Link.__deepcopy__ root cause, different attribute), #571 (general Robot/Link architecture direction -- this is a concrete instance of "what state does Link vs Robot own").
Background
PR #699 (fixing #638:
payload(0)wasn't a no-op) landed a real fix, but review surfaced three related design questions it deliberately left out of scope. Grouped here since they're all facets of the same underlying question: what doespayload()actually own, and does that model hold up once robots stop being simple serial chains?Confirmed: payload doesn't survive
copy()Robot.payload()now remembers a link's pre-payload m/r/I in a privatelink._payload_ownattribute sopayload(0)can restore it. That attribute is Robot-level bookkeeping, but it's stored on theLinkobject -- andLink.__deepcopy__(correctly, from Link's own point of view) only reconstructs the named dynamic parameters it knows about, so_payload_ownnever survives a copy. Reproduced directly:p560_copy.rne(...)afterpayload(0)is bit-for-bit identical to with the payload, not the bare-link baseline. No exception, no warning -- the copy's payload is permanently baked into what it now believes is the link's own mass. PR #699 documents this as a known limitation in its docstring, but "known" isn't "fixed," and the failure mode is silent wrong dynamics, not an error.This is the same root pattern as #577 (
Link.__deepcopy__silently drops coal collision geometry):Link.__deepcopy__only carries a fixed set of named parameters, so anything stored as an ad-hoc attribute onLink-- regardless of which class conceptually owns it -- disappears across a copy with no signal to the caller.Proper fix: move the "a payload is attached, here's what to subtract" bookkeeping onto
Robot/BaseRobotitself (e.g.self._payload: tuple[Link, float, NDArray, NDArray] | None), the same wayRobot.__deepcopy__/DHRobot.__deepcopy__already explicitly re-attach other instance-level state beyond the constructor kwargs (configs,ikine_a). That keepsLinkas a dumb parameter container and fixes the copy bug at the source, rather than teachingLink.__deepcopy__about a Robot-level concept smuggled onto it.payload()has no answer for multiple end-effectorsFor a robot with
len(ee_links) > 1, #699 falls back to the pre-existing behaviour of usinglinks[n - 1]-- whichever link happens to land last in the flattenedlinkslist. For a genuine tree that's arbitrary: it has no necessary physical relationship to either branch, and is fragile to construction/traversal order.The idiomatic fix is the
end=parameter pattern already used throughout the kinematics API (fkine,jacob0, etc., viaself._getlink(end, default)): givepayload()anend=parameter, defaulting toee_links[0]only when there's exactly one, and require it explicitly (raise, don't guess) when there's more than one.Robot.rne()'s branching support has no test coverageWhile investigating the above, confirmed empirically that
Robot.rne()does correctly handle a genuine kinematic tree (multiple independent moving-joint chains sharing a parent, not just static leaf links) -- built a synthetic two-armed robot and verified gravity-torque superposition holds and there's no cross-branch leakage. Thegroup_of_link_idx/backward-force-summation code inRobot.rne()clearly was written deliberately (extensive comments referencing #636/#483/#571), but there is currently no test in the repo that exercises it -- the existing "branched robot" tests (test_dotfile2,test_graph_multiple_end_effectors, YuMi'sfkine/jacob0tests) only touch kinematics/graph rendering, and #699's owntest_payload_several_end_effectorshas two static tip links off one joint (branching in link topology, not in DOF), so it never reaches the force-summation-at-shared-parent path.Worth a regression test using something like the synthetic two-armed robot above (or YuMi directly), independent of the payload/end= work.
Scope
Not blocking #699, which fixes a real, independently-valuable bug (#638) on its own. Related: #577 (same
Link.__deepcopy__root cause, different attribute), #571 (general Robot/Link architecture direction -- this is a concrete instance of "what state does Link vs Robot own").