[Bug]: Split narrowphase on GPU skips GJK when requires_grad=True, producing zero convex-convex contacts

[EntangledQuantum]

Bug Description

On GPU backends, Genesis enables split narrowphase whenever the scene contains non-box/plane convex–convex pairs (collider._use_split_narrowphase == True). In this path, contact0 correctly detects overlap and enqueues pairs into the GJK work queue, but _func_multicontact_run_detection never runs GJK (or differentiable GJK) when requires_grad=True.

The narrowphase path (func_convex_convex_contact) correctly branches on requires_grad and calls diff_gjk.func_gjk_contact plus func_add_diff_contact_input. The split multicontact path does not — GJK is guarded by if qd.static(not static_rigid_sim_config.requires_grad) with no else branch.

Symptoms:

• Broadphase works (n_broad_pairs increases when bodies overlap).
• Robot–object pairs appear in narrowphase_work_queues GJK queue.
• multicontact_gjk_state.is_col stays 0 and n_contacts == 0 for those pairs.
• Objects pass through each other (“ghost” collision) despite visual overlap.
• Same scene with requires_grad=False produces contacts normally.

Secondary issue: _func_multicontact_gjk_full writes contacts via func_set_contact but never calls func_add_diff_contact_input, so collider_state.diff_contact_input is never populated on the split path (even when forward contacts exist without grad). This breaks the differentiable backward pass (func_narrow_phase_diff_convex_vs_convex.grad).

Affected code:

        if qd.static(collider_static_config.ccd_algorithm != CCD_ALGORITHM_CODE.MJ_MPR):
            if use_gjk:
                if qd.static(not static_rigid_sim_config.requires_grad):
                    gjk.func_gjk_contact(...)
                    ...
                    used_gjk = True
                # MISSING: else branch with diff_gjk.func_gjk_contact(...) when requires_grad=True

Compare with the working monolithic path at narrowphase.py ~L1234–1299, which calls diff_gjk.func_gjk_contact and func_add_diff_contact_input when requires_grad=True.

Instrumentation note: Reading collider._gjk._gjk_state is misleading on split narrowphase — contact0 uses _contact0_gjk_state and multicontact uses _multicontact_gjk_state.

Steps to Reproduce

1. Use Genesis on a CUDA GPU with a scene that triggers split narrowphase (any convex non-box/plane pair, e.g. Franka finger BOX geoms vs USD/MESH peg).
2. Build a rigid scene with:
   • sim_options.requires_grad=True
   • rigid_options.use_gjk_collision=True
   • rigid_options.enable_multi_contact=True
   • rigid_options.box_box_detection=False (required when requires_grad=True)
3. Place a dynamic convex object (e.g. peg geom type MESH, idx 47) in forced overlap with robot finger collision geoms (e.g. geoms 11, 12).
4. Run collision detection and inspect split-narrowphase state after contact0, before multicontact pass 1:

collider = scene.sim.rigid_solver.collider
wq = collider._collider_state.narrowphase_work_queues
mc_gjk = collider._multicontact_gjk_state
After contact0: pairs are enqueued
gjk_size = int(wq.gjk_queue_size.to_numpy()[0])

print(list(zip(wq.gjk_i_ga.to_numpy()[:gjk_size], wq.gjk_i_gb.to_numpy()[:gjk_size])))
e.g. [(12, 47), (11, 47)]
collider.detection()

print("multicontact is_col:", (mc_gjk.is_col.to_numpy() == 1).sum())

print("n_contacts:", collider._collider_state.n_contacts.to_numpy()[0])

print("robot-peg contacts:", len(robot.get_contacts(with_entity=peg)["position"]))

1. Repeat step 4 with requires_grad=False on the same scene/pose.

Minimal repro script (in this repo):

scripts/utils/debug_split_narrowphase.py

conda run -n genesis python scripts/utils/debug_split_narrowphase.py

Observed results (RTX 5090, Genesis 1.0.0):

Metric	requires_grad=True	requires_grad=False
use_split_narrowphase	True	True
GJK queue (robot–peg)	(12,47), (11,47)	(12,47), (11,47)
multicontact_gjk.is_col true	0	2
n_contacts	0	2
robot.get_contacts(with_entity=peg)	0	2

Expected Behavior

When requires_grad=True on GPU with split narrowphase active:

1. _func_multicontact_run_detection should call diff_gjk.func_gjk_contact (mirroring func_convex_convex_contact) instead of skipping GJK entirely.
2. On collision, func_add_diff_contact_input should populate collider_state.diff_contact_input so forward contacts and backward gradients are consistent.
3. Robot–peg (and any similar convex–convex) pairs enqueued to the GJK queue should produce n_contacts > 0 and physical contact response, same as requires_grad=False.

Screenshots/Videos

Relevant log output

[Genesis] [02:11:23] [INFO] ╭───────────────────────────────────────────────╮
[Genesis] [02:11:23] [INFO] │┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈ Genesis ┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈┉┈│
[Genesis] [02:11:23] [INFO] ╰───────────────────────────────────────────────╯
[Genesis] [02:11:23] [INFO] Running on [NVIDIA GeForce RTX 5090] with backend gs.cuda. Device memory: 31.33 GB.
[Genesis] [02:11:23] [INFO] 🚀 Genesis initialized. 🔖 version: 1.0.0, 🎨 theme: dark, 🌱 seed: None, 🐛 debug: False, 📏 precision: 32, 🔥 performance: False, 💬 verbose: INFO
[Genesis] [02:11:25] [INFO] Scene <38e308c> created.
[Genesis] [02:11:25] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 0, uid: <52b7e81>, morph: <gs.morphs.Plane>, material: <gs.materials.Rigid>.
[Genesis] [02:11:25] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 1, uid: <406b17c>, morph: <gs.morphs.MJCF(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/genesis-world/genesis/assets/xml/franka_emika_panda/panda.xml')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:25] [WARNING] (MJCF) Approximating tendon by joint actuator for `finger_joint1`
[Genesis] [02:11:25] [WARNING] (MJCF) Approximating tendon by joint actuator for `finger_joint2`
[Genesis] [02:11:25] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 2, uid: <a19cefc>, morph: <gs.morphs.USD(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_hole_8mm.usd')(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_hole_8mm.usd', prim_path='None')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:25] [INFO] Convex hull is not accurate enough for collision detection (0.941). Falling back to more expensive convex decomposition (see FileMorph options).
[Genesis] [02:11:25] [INFO] Baked assets detected and used: /home/cognitron/.cache/genesis/usd/bake/592ee24f0b0034901c2eea5b7c0bb91e1e71cf9afbace4c4c846dedc679ceef0/factory_peg_8mm.usd
[Genesis] [02:11:25] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 3, uid: <68edd52>, morph: <gs.morphs.USD(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_peg_8mm.usd')(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_peg_8mm.usd', prim_path='None')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:25] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 4, uid: <f06c089>, morph: <gs.morphs.Mesh(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/genesis-world/genesis/assets/meshes/axis.obj')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:25] [INFO] Building scene <38e308c>...
[Genesis] [02:11:25] [INFO] Mass is not specified and collision geoms can not be found for link 'axis_obj_baselink'. Using visual geoms to compute inertial properties.
[Genesis] [02:11:25] [WARNING] Neutral robot position (qpos0) exceeds joint limits.
[Genesis] [02:11:26] [WARNING] Filtered out geometry pairs causing self-collision for the neutral configuration (qpos0): (6, 11), (7, 11), (14, 21), (15, 20), (16, 23), (17, 22). Consider tuning Morph option 'decompose_robot_error_threshold' or specify dedicated collision meshes. This behavior can be disabled by setting Morph option 'enable_neutral_collision=True'.
[Genesis] [02:11:26] [INFO] Compiling simulation kernels...
[Genesis] [02:11:27] [INFO] Building visualizer...

========================================================================
RUN: requires_grad=True mesh peg
========================================================================
  backend=2  requires_grad=True
  use_split_narrowphase=True
  peg geom idx=47  type=6
  n_broad_pairs=4
  n_contacts (collider)=0
  n_diff_contact_input.valid=0
  robot.get_contacts(with_entity=peg)=0
  peg total contacts=0
  WRONG collider._gjk._gjk_state is_col true count=0

  --- after_contact0_before_mc_pass1 ---
    gjk_queue_size: 2
    mpr_queue_size: 0
    gjk_pairs_peg: [(12, 47), (11, 47)]
    gjk_pairs_robot_peg: [(12, 47), (11, 47)]

  --- after_contact0_before_mc_pass1 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 0
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

  --- after_mc_pass1 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 0
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

  --- after_mc_pass2 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 0
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[Genesis] [02:11:29] [INFO] Scene <e380b9e> created.
[Genesis] [02:11:29] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 0, uid: <e1b14cd>, morph: <gs.morphs.Plane>, material: <gs.materials.Rigid>.
[Genesis] [02:11:29] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 1, uid: <21b7181>, morph: <gs.morphs.MJCF(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/genesis-world/genesis/assets/xml/franka_emika_panda/panda.xml')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:29] [WARNING] (MJCF) Approximating tendon by joint actuator for `finger_joint1`
[Genesis] [02:11:29] [WARNING] (MJCF) Approximating tendon by joint actuator for `finger_joint2`
[Genesis] [02:11:29] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 2, uid: <9cbd7ca>, morph: <gs.morphs.USD(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_hole_8mm.usd')(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_hole_8mm.usd', prim_path='None')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:29] [INFO] Convex hull is not accurate enough for collision detection (0.941). Falling back to more expensive convex decomposition (see FileMorph options).
[Genesis] [02:11:29] [INFO] Baked assets detected and used: /home/cognitron/.cache/genesis/usd/bake/592ee24f0b0034901c2eea5b7c0bb91e1e71cf9afbace4c4c846dedc679ceef0/factory_peg_8mm.usd
[Genesis] [02:11:29] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 3, uid: <9572e62>, morph: <gs.morphs.USD(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_peg_8mm.usd')(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/assests/factory_peg_8mm.usd', prim_path='None')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:29] [INFO] Adding <gs.engine.entities.RigidEntity>. idx: 4, uid: <afa6776>, morph: <gs.morphs.Mesh(file='/home/cognitron/Projects/Physics_Projects/Cognitron_MultiRobot_MVP/genesis-world/genesis/assets/meshes/axis.obj')>, material: <gs.materials.Rigid>.
[Genesis] [02:11:29] [INFO] Building scene <e380b9e>...
[Genesis] [02:11:29] [INFO] Mass is not specified and collision geoms can not be found for link 'axis_obj_baselink'. Using visual geoms to compute inertial properties.
[Genesis] [02:11:29] [WARNING] Neutral robot position (qpos0) exceeds joint limits.
[Genesis] [02:11:30] [WARNING] Filtered out geometry pairs causing self-collision for the neutral configuration (qpos0): (6, 11), (7, 11), (14, 21), (15, 20), (16, 23), (17, 22). Consider tuning Morph option 'decompose_robot_error_threshold' or specify dedicated collision meshes. This behavior can be disabled by setting Morph option 'enable_neutral_collision=True'.
[Genesis] [02:11:30] [INFO] Compiling simulation kernels...
[Genesis] [02:11:48] [INFO] Building visualizer...

========================================================================
RUN: requires_grad=False mesh peg
========================================================================
  backend=2  requires_grad=False
  use_split_narrowphase=True
  peg geom idx=47  type=6
  n_broad_pairs=5
  n_contacts (collider)=2
  n_diff_contact_input.valid=0
  robot.get_contacts(with_entity=peg)=2
  peg total contacts=2
  WRONG collider._gjk._gjk_state is_col true count=0

  --- after_contact0_before_mc_pass1 ---
    gjk_queue_size: 2
    mpr_queue_size: 0
    gjk_pairs_peg: [(12, 47), (11, 47)]
    gjk_pairs_robot_peg: [(12, 47), (11, 47)]

  --- after_contact0_before_mc_pass1 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 0
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

  --- after_mc_pass1 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 2
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

  --- after_mc_pass2 ---
    contact0_is_col_true: 0
    contact0_is_col_sample: [0]
    multicontact_is_col_true: 2
    multicontact_is_col_sample: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

========================================================================
COMPARISON (grad vs no-grad)
========================================================================
  n_contacts: grad=0  no-grad=2
  n_diff_contact_input_valid: grad=0  no-grad=0
  n_robot_peg_contacts: grad=0  no-grad=2
[Genesis] [02:11:50] [INFO] 💤 Exiting Genesis and caching compiled kernels...

Environment

Field	Value
OS	Ubuntu 24.04.4 LTS (kernel 6.17.0-29-generic)
GPU/CPU	NVIDIA GeForce RTX 5090
GPU driver	580.159.04
CUDA / CUDA toolkit	CUDA 13.3 (compiler build cuda_13.3.r13.3/compiler.37862127_0)
Genesis release	1.0.0
Commit ID	6007f263fd2c76586c65eb787de7bd0b0c67aee9 (v1.0.0-4-g6007f263)
Backend	gs.cuda

Release version or Commit ID

6007f26
v1.0.0-4-g6007f263

Additional Context

i am stuck on this for hours . i had a peg in the hole project with differtiable gradients enabled the robot and the peg (USD object) simply become ghost to each other even though they are colliding with the ground.

[EntangledQuantum]

Debugging script used
debug_split_narrowphase.py

[EntangledQuantum]

@duburcqa is there a easy fix this ? I am trying to do convex <--> convex contact with franks gripper holding a object with differentiablity enabled. but the contact/collsion simply won't work due to this issue above.

[duburcqa]

@EntangledQuantum It should be easy to fix! I will fix it asap.

[EntangledQuantum]

@duburcqa Thanks for the Fix. now the collision is getting detected when gradients are enabled . i have a question , is there a future propects of genesis into making stiff contact gradients from the simulators much better gradient wise ? for examples i found this simulator called DiffMJX which does that but unfortunatly doesn't support convex <> convex mesh contact.

2026-06-01.23-52-10.mp4

[duburcqa]

i have a question , is there a future propects of genesis into making stiff contact gradients from the simulators much better gradient wise ?

Could you elaborate? I'm not sure to understand. Note that there is a significant ongoing effort for dramatically improving the robustness of differentiable simulation: #2842