First draft of GPSR 2026

[adamgolding1]

Updating the GPRS Task
Changes proposed in this pull request:

• Update section in Procedure for Command execution to include giving the instructions consecutively or one-by-one
• Update section in Additional Rules for passing command through LLM to get similar command
• Add to Deus ex Machina QR code will not be available

Question
If the commands are executed consecutively, how is the arena re-arranged between tasks?

[LeroyR]

Question If the commands are executed consecutively, how is the arena re-arranged between tasks?

We should just generate 3 commands at once and setup the arena for that.

• Another Task is running parallel anyway -> we could increase the max duration.
• Better describe that the robot can request additional commands at any time (while at the instruction point) (even consecutively)
• Add rule for bonus points if commands are combined, planned/executed interleaved (example picks up an item while passing through a room, then finishes non manipulation task first before delivering the item at the target) and for stating the reason for its plan.
• We need to come up with more sensible tasks (update the generator)
  • Idea: fetch by object description like "bring me a CATEGROY that is Y" -> bring a drink that is red could you fetch me a sweet food item etc.
  • Find person doing task to tell him about appointment find michael to remind him about his doctors appointment. He should be in the apartment reading a book
    • Add penalty for talking to a person not matching the description?
      • Although if the robot searches and does not find the person, then completes other tasks, and then reports to me he could not find him and asks for direction or suggests talking to each found person it would be a good solution 😅
• Imho the rulebook needs to better describe what tasks are given (we should maybe generate X examples and include them in the rules)
• Should we remove the non-expert bonus if we LLM rephrase them anyway?
• How often is a command repeated? Do we generate different phrasing each time?

[ARTenshi]

As a reminder, the rulebook from 2023 and before already included:

"commands can be issued by a Naive Operator, i.e. a person from the audience with no background on robotics. The referee gives the command to the Naive Operator, who will then issue it to the robot (rephrasing is allowed). If the robot consistently fails to understand the naive operator (e.g. 3 times or more), teams can default to a custom operator."

"Commands for EGPSR are either Stage II tasks, complex commands requiring to perform chains of subtasks in sequence, or incomplete commands lacking relevant information to succeed."

Since rephrasing and multiple command attempts were already included in previous editions, I'm concerned this may not advance the competition's technical challenges.

A more challenging approach would be giving additional tasks while the robot is performing the first command. For example, the robot receives the first command at the instruction point; while executing it, a user calls the robot and requests a second task. The robot must then decide the order of execution—whether to finish the current task or pause it to handle the new request. The user (or an additional user) could then request a third task, requiring dynamic priority management.

Teams that can only handle one task at a time would complete their run after the first task (suitable for beginner teams), while more advanced teams could attempt to handle multiple concurrent requests for additional points.

Keeping the competition accessible to new teams remains important, but I think we should move forward on the challenges we proposed and not going back to what we have already done.

[LeroyR]

I assume "going back" => returning to previously solved tasks but while i could see not having enough forward movement i dont agree that this is "going back".

As a reminder, the rulebook from 2023 and before already included:

"commands can be issued by a Naive Operator, i.e. a person from the audience with no background on robotics. The referee gives the command to the Naive Operator, who will then issue it to the robot (rephrasing is allowed). If the robot consistently fails to understand the naive operator (e.g. 3 times or more), teams can default to a custom operator."

Which was Optional. "random/llm" rephrasing now mandatory. Previously you could generate every possible combination of commands for parsing.

"Commands for EGPSR are either Stage II tasks, complex commands requiring to perform chains of subtasks in sequence, or incomplete commands lacking relevant information to succeed."

Since rephrasing and multiple command attempts were already included in previous editions, I'm concerned this may not advance the competition's technical challenges.

Complex Subtasks and incomplete Task where not in the generator for the last 3 years. The incomplete commands (before that) never used or solved.
Rephrasing was Optional (now Mandatory).

A more challenging approach would be giving additional tasks while the robot is performing the first command. For example, the robot receives the first command at the instruction point; while executing it, a user calls the robot and requests a second task. The robot must then decide the order of execution—whether to finish the current task or pause it to handle the new request. The user (or an additional user) could then request a third task, requiring dynamic priority management.

Optional planning/merging of multiple tasks would be new.
Previous Optionals into Required is the step forward.
We also decided to revamp the generated commands which may increase the execution difficulty.
Other changes in the Rulebook are already massive which is why this task, in general, will not change too much.

Teams that can only handle one task at a time would complete their run after the first task (suitable for beginner teams), while more advanced teams could attempt to handle multiple concurrent requests for additional points.

Keeping the competition accessible to new teams remains important, but I think we should move forward on the challenges we proposed and not going back to what we have already done.

[ARTenshi]

Just to clarify, those "other changes" mostly concern the competition format, such as drastically reducing the number of tasks and allowing them to be repeated over several days. This simplification in the format should allow us to introduce more challenging tasks rather than "keeping" them as they are (with some improvements given the feedback from previous iterations).

[LeroyR]

Just to clarify, those "other changes" mostly concern the competition format, such as drastically reducing the number of tasks and allowing them to be repeated over several days. This simplification in the format should allow us to introduce more challenging tasks rather than "keeping" them as they are (with some improvements given the feedback from previous iterations).

While i personally agree (and also suggested more new tasks, multiuser, speech interruptions etc), the consensus was to not change e.g. receptionist substantially to avoid too many changes at once. I think adding a planning bonus here (give all commands instead of one at a time) while generating more interesting commands ensures this task is still interesting enough.

[adamgolding1]

Updates based on discussions from 6/10/2025

• remove optional goal - use non-expert operator
• referee will issue commands

Additional Rules and

• update Command Generator section
• update Custom Operators section

Score Sheet

• split main goal into recognising command, planning, execution
• removed optional goals

gpsr_2026_v_0_1.pdf

[adamgolding1]

Attached is a pdf version of GPSR and a list of commands from the current generator
GPSR_task.pdf
generated_commands_examples.pdf