VLA Action Inspect

The pack's robotics nodes give you plenty of ways to get action chunks - policy servers, JSON replay, safety gates - and exactly one way to actually look at them. VLA Action Inspect serializes a validated action trajectory and shows you one step of it in detail. It's a debugging tool, and it's unapologetically simple, which is what makes it useful.

The inputs are just two: actions (a VLA_ACTIONS handle from any policy node, the safety gate, or Actions From JSON) and step_index (which step of the trajectory to zoom into, default 0). Run it and you get four outputs:

• trajectory_json - the whole trajectory serialized as text, ready to save, diff, or paste into a script
• selected_step_json - just the step at step_index, with its per-dimension values
• horizon - how many steps the whole chunk spans
• action_dim - how many values per step

Those last two are the ones that catch real bugs. If a policy hands you a 32-step chunk but you told the safety gate to plan for 8, seeing horizon: 32 in plain INT form is the fastest way to realize the mismatch. If action_dim says 8 but your embodiment profile says 7, the trajectory is wrong before it ever reaches a motor.

The description carries an important guarantee: it "never writes to a robot transport." Nothing in this node can command hardware - it's pure read-and-serialize. So when you're staring at a trajectory that made the real robot do something terrifying, this is the safe node to reach for first: it will show you exactly what the numbers were, without any chance of replaying them.

Installing it

In the main pack - ComfyUI Manager (search "ComfyUI VLM nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes
python -m pip install -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements.txt

No extra robotics dependencies. This node runs with zero policy or robot SDKs installed, like all the non-client robotics nodes.

Common issues

• step_index out of range → you asked for step 12 of an 8-step chunk. horizon is your guide; pick a valid index.
• Action values look wrong → check whether they're in the embodiment's units (radians vs degrees, normalized vs raw). The node shows raw numbers; unit interpretation is the profile's job.
• Nothing to inspect → you fed it an empty or invalid VLA_ACTIONS; the node validates on the way in, so a hard error here means the upstream node emitted something malformed.

The natural workflow is a debugging chain: run the policy, connect its output to both the safety gate (for validation) and this node (for human eyes), then decide. When the trajectory looks right but the robot still misbehaves, the problem is downstream of ComfyUI - which is exactly the boundary this pack draws.