Rocky's Open-Source Build Thread (AI for Industry Challenge)

Competitions AI for Industry Challenge
12

March 14 Worklog Update

Big Win Today: Force Feedback is Working

Thanks to @jlamperez’s amazing writeup, I finally got force feedback working in my cheatcode teleop flow.

image
image1920×1000 637 KB

As you can see in the terminal output on the right side:


TFs found! Starting APPROACH phase.

[CheatCode] Force: 0.2N | Phase: APPROACH | z_off: 0.2000

[CheatCode] Force: 3.1N | Phase: APPROACH | z_off: 0.2000

[CheatCode] Force: 1.5N | Phase: APPROACH | z_off: 0.2000

[CheatCode] Force: 1.6N | Phase: APPROACH | z_off: 0.2000

[CheatCode] Force: 1.3N | Phase: APPROACH | z_off: 0.2000

Hover reached (err=0.0063m). Entering ALIGN phase.

[CheatCode] Force: 1.2N | Phase: ALIGN | z_off: 0.0500

[CheatCode] Force: 1.1N | Phase: ALIGN | z_off: 0.0500

Aligned! (xy=0.0024m, ang=0.000rad, dwell=1.7s). Starting INSERT.

[CheatCode] Force: 0.6N | Phase: INSERT | z_off: 0.0500

[CheatCode] Force: 0.6N | Phase: INSERT | z_off: 0.0100

[CheatCode] Force: 0.9N | Phase: INSERT | z_off: -0.0100

[CheatCode] Force: 1.1N | Phase: INSERT | z_off: -0.0100

[CheatCode] Force: 5.8N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 8.6N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 10.5N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.6N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.6N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.4N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 10.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.7N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.7N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.4N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 12.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.4N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.0N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.4N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.7N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 4.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 6.6N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 7.9N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 9.6N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 11.3N | Phase: INSERT | z_off: -0.0150

[CheatCode] Force: 13.1N | Phase: INSERT | z_off: -0.0150

(insertion completed)

Teleop loop time: 16.86ms (59 Hz))

The current state machine now incorporates force feedback, and this run gave me much better visibility into what is actually happening during insertion.

What I Changed Today

A) Added a full PROJECT_CONTEXT.md for future agents

I created a comprehensive context file at:

mystuff/for_agents/PROJECT_CONTEXT.md

It is 216 lines and includes:

  • competition overview

  • pipeline strategy

  • repository structure

  • aic_teleop.py architecture

  • scoring system

  • trial configurations

  • technical details

  • practical notes for AI agents

The goal is simple: if any future AI agent jumps into this repo, it should have enough context to contribute effectively without me re-explaining everything from scratch.

B) Rewrote AICCheatCodeTeleop (v2) in aic_teleop.py

I made major upgrades to the cheatcode teleop state machine:

  • Added a dedicated ALIGN phase

  • Dwell at 5 cm hover height for at least 1 second

  • Require XY error < 3 mm and angular error < 0.05 rad before INSERT

  • Slowed insertion descent from 0.07 m/s → 0.02 m/s

  • Added force-proportional insertion speed control

  • starts slowing at 5N

  • fully pauses at 15N

  • replaces old binary go/stop behavior

  • Improved recovery behavior

  • retreat only to 5 cm (not 20 cm)

  • return to ALIGN (not APPROACH)

  • maximum 3 retries

  • Added live XY correction during INSERT by continuously re-reading plug TF

  • Tightened phase transitions

  • dist_to_target < 0.01 (instead of the loose 0.2 value used before in recovery logic)

  • Added angular convergence gate before entering INSERT

  • Clamped max linear velocity during INSERT to 60% for gentler motion

  • Tuned gains:

  • kp_linear: 1.0 → 1.2

  • ki_linear: 0.15 → 0.2

  • kp_angular: 1.5 → 2.0

  • max_linear_vel: 0.1 → 0.08

C) Lowered retreat threshold from 18N to 10N

I originally wrote a recovery mode that retreats at 18N, but looking at the terminal evidence above, the robot can get badly stuck around 10-13N without ever reaching 18N.

So with the old threshold, it could keep pushing while stuck and just hope for the best.

New behavior chain is now:

  • 0-5N: full speed

  • 5-15N: linear slowdown

  • ≥15N: full stop

  • ≥10N for 0.5s: retreat and retry

This should make the behavior much safer and reduce the chance of force-penalty disasters.

Side Project Note

I have also been working on a parallel project I started a couple of months ago, but not consistently:

Tooling Note (Slate)

I also found a new coding tool called Slate. A lot of today’s changes came from working through one giant, unorganized prompt with it. I’m definitely not an expert in coding, engineering, or AI-agent workflows yet, but this tool exceeded my expectations in the first hour, so I wanted to share that.

image
image1920×1000 359 KB

Slate also pointed out that my current cheatcode teleop is using velocity control instead of position control. I haven’t fully verified that myself yet, so I can’t confirm 100% today, but I’ll dig into it tomorrow.

Overall, today was a solid step forward: better force-aware behavior, cleaner state transitions, and better project context for future iterations.