impulselabsRESEARCH ROADMAP
IMPULSELABS / THE RESEARCH DIRECTION

First, a shot.
Then, a bigger game.

Soccer combines movement, ball control and decisions. We begin with one focused task: shooting at goal in a physics simulation.

The longer-term goal is to explore how individual skills can develop into play. This roadmap describes research directions, not a schedule of guaranteed capabilities. The order and scope may change as we learn from the first experiment.

STAGE 01 / CURRENT FOCUS

Shooting at Goal Training setup in progress

Learn the movements and ball contact involved in a shot, then measure what happens to the ball. The initial task is deliberately narrow enough to examine through repeated attempts and comparable evaluations.

Research focus
Coordinated body movement, contact with the ball, shot direction and measured outcomes at the goal line.
Training process
Learn reference movements, continue with ball interaction, then evaluate saved versions separately. These are steps inside Stage 01, not separate soccer roadmap stages.
Evidence to examine
Goals and total attempts, goal-line error where recorded, evaluation conditions and checkpoint comparisons. Reward and goal rate are read separately.
Current state
The training package is assembled. A GPU run, convergence and connection of the new process to the dashboard have not yet been confirmed.
An archive is already available.

The dashboard contains 11 checkpoint entries and recordings from an earlier custom experiment. Its completed run and difficulty levels are not evidence that the new pipeline is running or that Stage 01 has been completed.

Explore the recorded experiment ↗
IMMEDIATE WORK / WITHIN STAGE 01

Make the new process observable.

Before claiming results for the new pipeline, the project needs to establish what is actually running and connect the observations to their source. The following items describe work to verify, not completed milestones.

  1. Validate a training run

    Confirm that the configured training process runs on the intended compute setup. A successful start is separate from convergence or improvement.

  2. Connect the dashboard

    Publish actual frames, run status and events. Tie each displayed measurement to its episode, checkpoint and simulation time.

  3. Save and evaluate versions

    Record checkpoint identity and evaluate on defined conditions, keeping training attempts, validation checks and held-out results distinct.

  4. Publish inspectable evidence

    Expose available measurements, test conditions and recordings. Show failures, missing data and stopped runs as clearly as successes.

There is no claimed completion percentage, launch date or automatic guarantee of improvement.

FUTURE DIRECTION

Ball Control Planned

Explore receiving, carrying and maintaining control of the ball. A single contact becomes a sequence of interactions: the robot must keep the ball available for a subsequent action.

The question
How can learned movement support repeated ball interaction rather than one isolated shot?
What needs to be defined
A dedicated task, its start and termination conditions, evaluation measurements and a baseline. None of these are presented as implemented for this direction yet.
What is not established
Reliable receiving, dribbling or possession. The shooting archive does not demonstrate these capabilities.
FUTURE DIRECTION

Passing Planned

Explore directed passes and interaction with a teammate. The research focus expands from producing a shot toward coordinating ball movement with another participant.

The question
How could an individual ball skill develop into an interaction that supports a teammate’s next action?
What needs to be defined
The partner setup, task conditions and evaluation protocol. Passing requires its own experiment and measurements rather than reusing shooting accuracy as proof.
What is not established
Successful passing, teammate coordination or cooperative play in the current experiment.
LONGER-TERM DIRECTION

Game Situations Research direction

Explore choosing actions with respect to the ball, teammates and opponents. This is a broader question than executing a movement: it concerns when and why an action should be taken.

The question
How might learned movement and ball skills be combined with decisions in a changing game situation?
What needs to be defined
Scenario rules, participants, observations and a suitable evaluation protocol. These depend on research findings and are not fixed commitments.
What is not established
Full-match play, strategy, autonomous teamwork or performance against opponents. These are not results of the current shooting task.

Progress in one focused simulation task does not by itself establish general soccer ability or performance on physical hardware.

HOW THE ROADMAP EVOLVES

Let the results set the next question.

A future direction should be described through its own task, measurements and history once it becomes an actual experiment. We will not turn a successful highlight into a claim that a research stage is complete.

  • Use comparable repeated evaluations to assess changes between saved versions.
  • Keep the evaluation conditions and sample counts visible alongside rates.
  • Distinguish goals from balance, reward from accuracy, and a new checkpoint from a better one.
  • Mark planned directions as planned until there is evidence for a stronger claim.
  • Update the research scope when results suggest a different next step.
What “planned” means here

A direction we intend to investigate. It is not a shipping promise, a readiness score or an assigned date. Research can improve, stall or reveal that a different approach is needed.