Joint velocities exceed limits immediately after reset in torque control (Franka)

[moribots]

Hi all,

Setup

• Isaac Sim 5.0, Isaac Lab 2.2 (feature/isaacsim_5_0)
• OS/GPU/driver: Ubuntu 22, RTX 5080, 577.144
• Task: custom Franka reach
• Control: torque only (JointEffortAction), PD off (stiffness=0, damping=0)
• Limits set with *_limit_sim fields

Symptoms

On every reset(), terminations.joint_vel_out_of_limit trips because some joints have |dq| > limit. This happens before any action is applied and with effort=0 (i.e. falling due to gravity).

Actuators

actuators={
    "panda_shoulder": ImplicitActuatorCfg(
        joint_names_expr=["panda_joint[1-4]"],
        effort_limit=87.0,  effort_limit_sim=87.0,
        velocity_limit=2.1750, velocity_limit_sim=2.1750,
        stiffness=0.0, damping=0.0,
    ),
    "panda_forearm": ImplicitActuatorCfg(
        joint_names_expr=["panda_joint[5-7]"],
        effort_limit=12.0,  effort_limit_sim=12.0,
        velocity_limit=2.6100, velocity_limit_sim=2.6100,
        stiffness=0.0, damping=0.0,
    ),
}

Reset sequence (based on docs)

# Event order: place props -> zero-vel reset; no commands on reset frame

# 1) place goal/shelf (kinematic USDs), then:
asset = env.scene["robot"]
env_ids = torch.arange(env.num_envs, device=env.device)

default_root = asset.data.default_root_state[env_ids].clone()
if hasattr(env.scene, "env_origins") and env.scene.env_origins is not None:
    default_root[:, 0:3] = env.scene.env_origins[env_ids]

q_default = asset.data.default_joint_pos[env_ids].clone()
dq_zeros = torch.zeros_like(asset.data.default_joint_vel[env_ids])

asset.write_root_pose_to_sim(default_root[:, :7], env_ids=env_ids)
asset.write_root_velocity_to_sim(torch.zeros((env_ids.numel(), 6), device=env.device, dtype=default_root.dtype), env_ids=env_ids)
asset.write_joint_state_to_sim(q_default, dq_zeros, env_ids=env_ids)
asset.reset(env_ids=env_ids)

What I verified

• No set_*_target or action writes on the reset frame.
• Blank environment, no objects besides the Frnaka
• Articulation root: solver_position_iteration_count=8, solver_velocity_iteration_count=4 (also tried 0), sleep_threshold=0.0, stabilization_threshold=0.001.
• soft_joint_vel_limits match Franka specs.

Observed

Immediately after reset(), some joints report |dq| ≳ limits (e.g., j2 ≈ 2.19 rad/s vs 2.175).

Questions

• Is it expected that velocity “limits” are not enforced at reset time for articulations, i.e., contacts or depenetration can generate dq > limit even with effort=0?
• What is the recommended, supported way in Isaac Lab to guarantee dq=0 after reset() under pure torque control?
• Is there any additional API call or articulation/root property needed to ensure no contact impulses affect the joints at reset?

[ooctipus]

if gravity is turned off does this still happen? If it doesn't then
asset.write_root_velocity_to_sim(torch.zeros((env_ids.numel(), 6), device=env.device, dtype=default_root.dtype), env_ids=env_ids)
should be correctly setting the velocity to 0.

So the if the issue is that joint position can't hold after immediately due to gravity,

maybe you could tune down sim_dt?

you could try set actuator_cfg.velocity_limit_sim to a desirable number so joint speed can't exceed that limit at simulation level.

or you could define a more loose termination that is not too strict.

If you application is RL, you could make velocity out of limit a penalty rewterm rather than termination term

[moribots]

If I turn off gravity and apply zero effort actions, the arm doesn't move at all, so the issue doesn't occur.

I will try your suggestions, thank you!

[moribots]

Yes this is for an RL application. I suppose I could create a curriculum that starts with having joint violations be a penalty, and eventually switch them to a termination condition once a certain success threshold has been met. Thanks!