Skip to content

Field days 2026-10-27..29: real-robot restarts, robot profiles, dashboard demos - #152

Merged
isaac0804 merged 54 commits into
mainfrom
field-day-20261027
Nov 2, 2026
Merged

isaac0804 merged 54 commits into
mainfrom
field-day-20261027

Conversation

@isaac0804

@isaac0804 isaac0804 commented Oct 28, 2026 •

Copy link
Copy Markdown
Contributor

Fixes and tools from the first real-robot field day (2026-10-27) and the analysis of its logs.

Referee restarts on real robots

  • Box rules come from the referee profile. Field.defense_area_rules follows defense_area.enabled, with no separate flag; the lab profiles have it off, so robots there ignore the boxes.
  • Placement next to the enemy box may enter its keep-out. Before, the placer was held 0.25 m short of the ball.
  • The placer approaching the ball:
    • it goes round the ball with 5 cm clearance, facing it (before, about 1 cm, facing the target);
    • it creeps onto the ball from the approach point;
    • it carries the ball at a capped speed;
    • it meets a rolling ball on its path instead of chasing it facing away.
  • Free-kick and penalty approaches use the same detour.
  • Clearing from the ball keeps off other robots. A robot sent clear of the ball (STOP, placements) now keeps 2 robot radii + 10 cm from every other robot, moving on round the ball if its spot is taken.
    • Source: the 38 logs from 2026-10-27 show 3 real restart collisions. The rest of the flagged contacts were vision errors or ball contests in play.
    • In the STOP one, our robot's target landed 0.23 m from their keeper; it drove into the keeper and pushed it 0.3 m.

Skills

  • The kick gate checks that the line along the robot's heading would score (heading_on_target).
  • The keeper stands on the bisector of the shot angle.
  • Clearances may go within 0.2 rad of upfield.
  • No more turning forever on the ball. A gap too narrow to aim into (narrower than twice the aim error) now counts as no shot, so the shooter moves for a better one (shot_heading_reachable, in _score_goal, LeadAndSupportTactic and GiveAndGoTactic).

Dashboard Live tab

  • Who drives each robot. During a restart or HALT a robot's row shows its job, e.g. referee · ball placement · placer, referee · stop · keeping clear or halted, and its ring on the field turns grey. Before, a placer showed its old tactic and looked like it was going to the ball.
  • Each robot's step shows on a line under its tactic (shoot: turning to shoot · go_to_ball.approach: direct). A run with a dashboard or a game log keeps a match log in memory for this.
  • The game log notes each change of a robot's row, so a field log says what each robot was trying to do.

Real-robot plumbing

  • Ball-sensor flicker grace of 0.3 s, in every mode.
  • Robot feedback timeout cut to 0.2 s.
  • One tick of tolerance after a control-loop stall before vision counts as lost.
  • The dashboard shows the ball sensor.

Tools and examples

  • pixi run calibrate: a terminal CLI that measures one robot (driving, turning, dribbling, kicking).
  • Pass-receive and keeper-feed demos.
  • examples/save_and_counter.py: a 1v1 with a goal each, where a keeper that saves takes the ball and counterattacks. In rsim with the real kick speed: 48 goals and 12 counterattacks in 52 rounds.
  • examples/passing_rally.py: two robots pass their way up the field in two lanes and shoot into the empty goal (PassReceiveTactic gains overridable role/open_spot/target/partner_set). In rsim with the real kick speed: 23 of 23 rallies scored.
  • Delay prediction for real robots: move() plans from our robots' pose 79 ms ahead (PoseAhead, RobotParams.CONTROL_LATENCY_S; 0 in the sims, so rsim is unchanged and the motion benchmark is identical). In rsim given the calibrated robot's delay: drift while driving and turning 30 → 7 cm, benchmark collisions 7 → 3.
  • trajsample plans with the real robot's weaker sideways acceleration (RobotParams.MAX_SIDEWAYS_ACCELERATION: real 1.8, None in the sims, so sim results are unchanged). In rsim given the calibrated robot, every scenario × 5 noise seeds: fpp 27/145 failed, trajsample 13, with the limit 12, so the limit itself barely matters. Real robots stay on fpp (docs/motion_planning_comparison.md).
  • Renamed examples. The real_ prefix is gone, since StrategyRunner --mode picks real, grSim or rsim. real_attacker_defender.py is now shooting_drill.py.
  • Quieter start. The vision stream is off by default (enable_vision_stream=True brings it back); waiting for vision prints one line when what it sees changes; pygame's banner and AVX2 note are hidden.

Robot profiles and calibration (2026-10-28/29 field days)

  • A robot is its radio. Each one's measured numbers live in utama_core/config/robots/radio_<n>.json (RobotProfile): loop delay, speed gain, acceleration, vision noise, tag offset, kick speed, bend, bias and spread, carry caps.
  • pixi run calibrate writes the profile in one command. The robot fetches the ball itself for a carry ladder and 10 charged kicks. --check tries a new profile against the committed one on the field in alternating blocks.
  • Vision corrected by where each robot's tag sits on its lid (calibrate measures it).
  • Passes and shots aimed against each robot's kick bend and bias (the ball's offset on the dribbler bends a kick: radio 0 4.2°/cm, radio 1 1.2°/cm).
  • Carry held to 0.3 m/s and 1 rad/s on real robots, where they lost the ball; each profile can set its own caps.
  • Real-robot tuning: 1.5 m/s and 6 rad/s, a stiffer turn PID, a 0.3 rad/s minimum turn, 0.07 rad aim tolerance before passing.
  • pixi run find-robots pairs each radio with the lid it wears: each radio that answers turns on the spot, and the lid that turned is its. The pairing is saved, and the examples check it on start: they find the robots again only if a saved lid is out of view.
  • Kickers kept charged. A real robot fires an empty kick when 25 s have passed since its last one, with the ball well away.
  • rsim kicks as calibrated. A run given profiles kicks those robots at their measured speed, bend, bias and spread. The evaluation passes none, so its matches keep the sim's kick.
  • Planner A/B field drill (examples/planner_ab.py) and a close-call scan of field logs.

Field days from one dashboard

  • pixi run demos: one dashboard stays up. The Live tab's Demo panel switches between the shooting drill, pass-receive, save-and-counter, score goal and teleop while halted:
    • the robots stop and the log is saved;
    • the radio and vision are released, and the next demo starts;
    • the page reloads by itself, and the last demo's panels go with it.
  • Settings in the panel for the next run: first robot, camera half, rules.
  • Field check (pixi run field-check: cameras, packet rates, ball and lids in view, then find-robots) from a button, with its lines shown under the panel; the demo starts again after.
  • Teleop on the page:
    • WASD/QE drive the chosen robot after Force Start, B kicks, F works the dribbler;
    • it stops on key release, on Halt, or 0.3 s after the page goes quiet.
  • Radio health: the Robots panel shows each robot's share of its commands answered over the last 2 s, red under 80%.
  • StrategyRunner.close now frees the radio's serial port and ends its vision thread, so another runner in the same process can take them.

Docs

  • docs/field_log_replay.md: design for an index / window / rerun tool for field logs.
  • docs/roadmap.md item 10: the calibration to-do list for a full robot profile.
  • New rows in the bug table in docs/STRATEGY_DEVELOPMENT.md.

Not done

  • Demo switching on real hardware is tested in rsim only. Whether the radio port reopens cleanly after a switch needs a field check.
  • Kick strength varies (radio 1 weak on 2026-10-28), with many possible causes, the battery among them. Not chased.
  • Robots meeting head-on (their placer running in while ours clears away): needs a planner that accounts for the 79 ms delay and sideways acceleration of ~1.8 m/s². Listed in the roadmap.
  • No round-robin. The restart, kick and shot-gap changes alter sim results; it was skipped on purpose.

Tests

  • pixi run test --headless: 13399 passed, 83 xfailed.
  • pixi run lint: clean.
  • Each fix has a regression test that fails without it.

isaac0804 and others added 9 commits October 27, 2026 22:44
real_attacker_defender.py --calibrate expects two robots, and the field check halts on the
one never seen; with a single robot the scripted moves now run here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Runs real_score_goal.py --calibrate; real_robot_setup.md gains the step and tools.md the task.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…radio 0

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Box rules follow the referee profile's defense_area.enabled (Field.defense_area_rules).
- Placement near the enemy box may enter its keep-out.
- The placer goes round the ball with 5 cm clearance facing it, creeps onto it, carries it
  at a capped speed and meets a rolling ball on its path.
- Free-kick and penalty approaches use the same detour.
- Robots cleared from the ball keep 2 robot radii + 10 cm from every other robot
  (STOP collision, field day 2026-10-27).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fc3fwMFr77jDeXbRUnd54Y
…ooser clearance aim

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fc3fwMFr77jDeXbRUnd54Y
… dashboard ball sensor

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fc3fwMFr77jDeXbRUnd54Y
@isaac0804 isaac0804 added the release:minor Minor changes to main label Oct 28, 2026
isaac0804 and others added 20 commits October 28, 2026 15:51
…m into

find_best_shot returned any gap, but the shot gate (heading_on_target) only kicks
when the line stays clear across the whole aim error. In a gap narrower than that
the shooter turned on the ball forever. shot_heading_reachable keeps a gap only
if it is clear with twice the aim error, so a robot anywhere within the aim error
of the heading gets a clear line; otherwise the no-shot reposition runs.

The three shooting branches also trace each step (fetching the ball, turning to
shoot, kick, no clear shot) through the new TickContext.trace, for the Live tab.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the game log

On the field day a ball placement looked like a robot going to the ball: through a
restart the Tactics panel kept showing the slot's tactic. While a restart or HALT
drives a robot its row now says so, with its job (referee · ball placement · placer,
referee · stop · keeping clear, our free kick · kicker, halted), from
Strategy.referee_control and RefereeOverride.jobs; its field ring turns grey.

Each row also gets the steps its tactic and skills traced this tick (shoot: turning
to shoot · go_to_ball.approach: direct). A run with a dashboard or a game log keeps
a match log in memory for this, and the game log gets a note whenever a robot's row
changes, so a field log says what each robot was trying to do.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on wait

- enable_vision_stream defaults to False: the Live tab draws the same field, and the
  stream was one more port and URL at every start.
- Waiting for vision prints one line, and only when what vision sees changes, not a
  five-line block every 3 s.
- pixi's activation env hides pygame's banner and its AVX2 note.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
save_and_counter.py: a 1v1 with a goal each, where a keeper that saves takes the
ball and counterattacks. Each robot reads from the shared world every tick whether
it keeps or attacks; the attacker carries the ball to 0.7 m and shoots.

The examples run in any mode through StrategyRunner (--mode real|grsim|rsim), so the
real_ prefix said nothing: real_attacker_defender is now shooting_drill (the keeper
feeds the ball back, so the drill loops), and the others drop the prefix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pass_receive.py's pair, going somewhere: the receiver gets open ADVANCE_M
upfield in the lane across from the ball, the passer waits until it is
there (otherwise the rally went back and forth), and the robot on the ball
shoots at the empty goal from SHOOT_RANGE_M while the partner trails.
PassReceiveTactic gains role/open_spot/target/partner_set to override.

rsim with the real 2.1 m/s kick: 23 of 23 rallies scored. The one stall
seen earlier was a ball within the planner's wall clearance of the end
line; the docstring says to start 0.5 m out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nd lands

Real robots act on a command about 79 ms after vision saw them
(2026-10-27 calibration). move() now plans from our robots' pose that far
ahead (PoseAhead: velocity and the turn rate between frames) and turns the
world velocity into robot-frame commands with the predicted heading.
RobotParams.CONTROL_LATENCY_S: real 0.079, sims 0, so rsim is unchanged
(motion benchmark identical).

rsim given the calibrated robot's delay: driving while turning half a
turn, drift off the line 30 -> 7 cm; benchmark collisions 7 -> 3.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eration

The calibrated real robot accelerates sideways at ~1.8 m/s^2 against 4-5
forward. RobotParams.MAX_SIDEWAYS_ACCELERATION (real 1.8, None in the
sims so their results are unchanged) slows each trajsample candidate by
its share across the robot's heading, so a sidestep the robot cannot make
in time is seen to collide.

rsim given the calibrated robot (delay, gains, per-axis acceleration),
5 close scenarios x 10 noise seeds: fpp 25/50 collisions, trajsample
10/50, with the sideways limit 9/50 and full suite 29/29 (fpp 24/29),
about 30% slower in crowds. fpp plans paths, not motions, so it cannot
use the limit; docs/motion_planning_comparison.md has the table.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eeded

The earlier table came from 5 scenarios chosen because fpp failed them and
from single runs (29/29). Over all 29 scenarios x 5 noise seeds: fpp 27
of 145 failed, trajsample 13, trajsample with the sideways limit 12. The
sideways limit barely matters; trajsample's advantage is the planner.
Adds why fpp suffers (a path, not a trajectory, is checked).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ently

Shooting drill on the real robots (2026-10-28): with the keeper leaving only
narrow gaps, every pickup went to "no clear shot, moving for one", which
moved at full speed with a turn to face the goal: up to 0.9 m/s backwards
and 3.9 rad/s, and the ball was off the dribbler within 0.2 s. The branch
also ran without the ball, so a robot that had just lost it strafed away.

_score_goal now fetches the ball first. The no-shot strafe in all three
shooting tactics uses move_utils.carry: move with the dribbler, scaled down
to CARRY_SPEED 0.5 m/s and CARRY_TURN 2 rad/s together (now in
physical_constants, shared with the dribbler-less profile).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o the goal before sliding

Field runs 2026-10-28: the shooting drill, save_and_counter and
dribble_and_shoot lost the ball within 0.2 s of every pickup, reversing and
sliding at up to 1 m/s while turning at 4 rad/s (the planner's limits).
4df81a1's 0.5 m/s / 2 rad/s cap on the shot strafe still lost it. The runs
that keep the ball (pass_receive, passing_rally) held it at <= 0.1 m/s and
<= 1 rad/s.

- RobotParams.CARRY_SPEED/CARRY_TURN (real 0.3, 1.0; None in the sims, so
  sim results are unchanged): move() scales a dribbling command down to both,
  together, while the robot has the ball. Covers every tactic and the pivot
  in turn_on_spot. Replaces 4df81a1's carry() helper.
- move_for_a_shot (shared by the three shooting tactics): with no clear shot
  a robot facing away from the goal pivots on the ball to face it first,
  then slides; turning while sliding is what lost the ball.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The real robots' motors can't turn slowly enough to close a smaller error, so a
pass could wait on a tiny one (field day 2026-10-28). ORIENTATION_TOLERANCE_RAD
0.05 -> 0.06. SHOT_AIM_ERROR_RAD stays 0.05: at 0.06 the shot gate's
twice-the-error line no longer fits a 1 m goal from beyond ~4 m, and the shot
still checks the robot's actual heading. The off-centre pass tests use 4 cm
(was 3.6) so their miss stays outside the wider tolerance.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ce of

FPP kept a field wall between robot and target unless the target was within
OBSTACLE_CLEARANCE (0.27 m) of it. A robot inside that clearance, heading for a
target just outside it, got a 0.7 m detour away from the target. On the
2.25 m-wide lab field (2026-10-28, pass_receive, blue 0) the passer lined up
on a ball 0.30 m from the touchline, drove onto it, was sent back, and
repeated every 1.4 s. A straight run between two points on the same side of a
wall never gets nearer it than the nearer end, so dropping it is safe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… slowed

0293441 slows a robot with the ball after calculate(), but the PID
controllers' acceleration limiters kept their own faster output (~1 m/s, 4
rad/s) as the value to ramp from. When yellow 1 lost the ball mid-dribble
(dribble_and_shoot, 2026-10-28), go_to_ball ramped down from that phantom
speed: 0.9 m/s sideways, away from the ball. MotionController.sent() tells
the controller what was actually sent; FPP and PID set both limiters from
it (no-op elsewhere; trajsample plans from the measured velocity).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Real kicks leave the heading towards the side the ball sits on the dribbler:
over 98 kicks (2026-10-27/28) yellow 1 bent 4.7 deg and blue 0 2.1 deg per cm
off-centre, no constant term. RobotProfile.kick_bend_deg_per_cm sets it per
robot; StrategyRunner hands it to the motion controller on real robots only, so
the sims are unchanged. Passes, shots and the shot gate aim that much the
other way (kick_bend).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… kicks

A kicker left unused runs flat, so the first measured kick barely moved the
ball. The ball part of the calibration script now starts with an empty kick.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e ball

A keeper that saved the ball facing its own goal drove off towards the other
one while turning, sideways, and lost the ball. Facing more than 0.3 rad off
the goal, it now pivots on the ball first, as move_for_a_shot does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At STOP a keeper between the ball and its goal line was inside the 0.8 m
keep-out; the planner's push-out ray from the ball through the robot ran
0.5 m past the goal line and the robot drove into its goal (lab field,
2026-10-28). The ray is now kept inside the field, and where that would
bring the robot closer to the ball it follows the referee step's target.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seen on the dribbler's edge but never on the ball sensor, a robot pivoted
towards the goal, turning away from the ball and pushing it off, then went
back for it, round after round into a corner (save_and_counter, 2026-10-28).
Without the sensor it now goes on fetching the ball before it turns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…enced from its output

A still robot vision saw in one frame of five swung +-0.4 m round where it
stood (2026-10-28 passing rally): each snap back onto a detection read as up
to 3 m/s, which the filter coasted on through the next gap. Each axis is now
a constant-velocity Kalman filter with velocity in its state.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
isaac0804 and others added 13 commits October 28, 2026 21:57
…w's kick bias

2026-10-28 passing rallies: all 23 passes that rolled out were yellow's, 22
anticlockwise of the receiver by 5-17 deg. Asked 0.1 rad/s on its last turn,
the passer stood still short of its aim and kicked at the 0.07 rad
tolerance's edge on the side it turned from (+3.5-4.7 deg). move() now turns
a real robot at least RobotParams.MIN_TURN while it is off its orientation,
so the tolerance goes back to 0.05, as the sims had it.

Its kicks also bent more than modelled: 5.0 deg/cm plus a 2.3 deg
anticlockwise bias with the ball centred (51 kicks, rms 3.8).
RobotProfile.kick_bias_deg is aimed against with the bend. The yellow robot
reads as vision id 0 now.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on it

Turns on the spot both ways at two rates, drives straight forward and back,
then has the operator press the robot against a straight edge. Prints the
tag's offset from the robot's centre (circle fit to the spins), the drive
direction off the tag's heading, and the tag's heading offset against the
edge; writes <log>.tag.json. Old calibration logs: blue 0's tag 1.1 cm
behind and 0.4 cm right of its centre, drives 3-5 deg off its heading;
yellow's within 0.5 cm and 1 deg. calibrate_robot.py split into
build_runner/play/parse_args to share them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…re_goal plays from the right

The rally tests pin the attacked goal to the right, as their cases are drawn.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…es 0.5 s without vision

Team decision after the 2026-10-28 field runs: real robots raised from 1 m/s, 4 rad/s and
50 rad/s^2 to 1.5, 6 and 100. Turn kp 1.4 -> 3.5 (90 deg turns crawled in). A robot lost to
vision for 0.2 s halted a run; the Kalman filter coasts it, so lost/silent limits are 0.5 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ive orbits faster

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on it

calibrate-tag measured yellow 0's tag 3.5 deg clockwise of the robot; nothing used it.
RobotProfile.tag_forward_m/tag_left_m/tag_heading_deg (calibrate-tag now prints them)
move each detection from the tag to its robot in PositionRefiner, real mode only.

Yellow's profile takes its tag; refitted with the heading corrected, its 63 kicks of
2026-10-28 give 4.4 deg/cm through zero (4.2 +1.2 deg with a bias, rms 5.8 vs 5.9), so
kick_bias_deg 2.3 is dropped: it was mostly the tag.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…10 kicks

The ball part of pixi run calibrate asked the operator to put the ball on the dribbler for
every move and kick, and one ball lost spoiled the rest of the fixed dribble script.
examples/ball_trials.py (BallTrialsTactic) fetches the ball lined up for each trial
(pass_receive.take_ball_facing, factored out), carries it 0.5 m each way and half a turn
each way at rising speeds until it loses it, then kicks 10 times along the field; it asks
only for the ball back when it is out of reach. The ladder prints the fastest carry each
way: what CARRY_SPEED/CARRY_TURN may be.

robot_calibration fits the kick bend (deg per cm off-centre, through zero) and corrects a
log by a robot's tag (FieldLog.correct_tags, --tag); calibrate uses the robot's profile
(shooting_drill.ROBOTS by radio ID). Yellow's bend 4.8 deg/cm from that fit over its
2026-10-28 kicks, tag corrected. The fixed BALL_SCRIPT is gone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…logs

examples/planner_ab.py: two robots swap ends head on and cross, in blocks alternating
fpp and trajsample (A B A B), recording per leg the nearest approach and the time, a
game log per block, and a summary per planner. rsim rehearsal, 16 legs each: no
contact either way.

Calibration and the drill now run under a referee with every rule off
(shooting_drill.no_rules_referee): without one the defense-area keep-out held robots
0.57 m off either box on the lab field (fpp stopped 0.34 m short of its spot).

analysis/close_calls.py: where two driven robots came within 0.25 m in field logs,
closing speed and referee command. 2026-10-27/28: 35 close calls, 8 contacts; 30 in
open play (14 in save_and_counter's 1v1), 5 under referee commands, all before the
restart fixes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…k tries it on the field

pixi run calibrate now writes utama_core/config/robots/radio_<n>.json, which every
example reads (load_profile): the tag offset from its turns on the spot, the kick bend
and bias fitted together (with standard errors), and carry caps from the ladder's
slowest direction, applied per robot by move() (RobotProfile.carry_speed/carry_turn).
No numbers copied by hand.

calibrate-tag is gone. Its straight-edge step measured the tag's heading, which the
kick bias fitted against the tag's heading already takes in; vision now moves only the
position. Seeded from 2026-10-28's kicks: radio 0 4.2 deg/cm +2.7 deg (55 kicks),
radio 1 1.2 deg/cm +2.8 deg (84 kicks).

--check alternates the committed and the new profile in blocks: carries at each
profile's caps and kicks aimed at a target against its bend and bias, printing hold
rates and the misses per profile.

Every kick in the ball trials comes charged: within 30 s of the kicker's last kick,
else the robot lets go and kicks once with no ball first; at least 1.2 s apart.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…les and calibrate use it

pixi run find-robots sends each radio ID a still command; each one that answers is
turned about 70 degrees on the spot and back, and the pattern that turned with it is its
lid. It prints radio, lid, whether the robot's profile was calibrated on that lid (the
tag offset is the lid's), silent radios and patterns no radio turned, and saves the
pairing on this machine (replays/field/robots_found.json).

A robot is its radio: the examples pick robots by radio (RADIO_0, RADIO_1) and on real
robots take each one's colour and pattern from the last find (shooting_drill.on_field),
so no vision ID or colour constant is edited by hand. pixi run calibrate runs the find
itself when not given both --robot and --radio, and takes --radio, the only robot, or
asks which.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… rsim kicks as calibrated

- Kicker charge: on real robots a robot that has not kicked for 25 s fires an empty kick when
  the ball is not on it or within 30 cm, never in a HALT (kicker_charge.py, applied in
  AbstractStrategy.step for controllers that keep kickers charged).
- Lids on start: examples and calibrate --radio watch vision 0.5 s; the saved pairing is used
  while each robot's lid is in view, else those robots are found again (find_robots.confirm).
- pixi run field-check: vision (cameras, packets/s, field, ball, lids), then find-robots with
  each robot's share of commands answered and its profile, as ok / !! lines.
- rsim kicks a robot given its profile as measured: kick_speed (+0.26 m/s rsim contact loss),
  bias + bend x ball offset, scatter kick_spread_deg (real_kick.py); aim compensation and carry
  caps now apply in rsim too. No profiles (the evaluation): the sim's kick, unchanged.
- RobotProfile.kick_speed / kick_spread_deg, filled by calibrate from the kick fit; committed
  profiles from 2026-10-28's kicks (yellow 0 2.19 m/s, rms 5.9; blue 0 1.31, rms 9.1).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One process keeps the dashboard up; the Live tab's Demo panel switches between the shooting
drill, pass-receive, save-and-counter and score goal while halted. The running demo closes
(robots stopped, log saved), the next is built on the same dashboard, and the page reloads by
itself. StrategyRunner.close now frees the radio's serial port and ends the vision thread
(its socket closed), so the next runner in the process can take them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n the Live tab

The Demo panel of `pixi run demos` now holds the settings changed most between runs (first
robot, camera half, rules; applied from the next Run), a Field check button on real robots
(tools/field_check.py's check with the radio freed, its lines under the panel, then the same
demo again) and a teleop demo (examples/teleop.py): WASD/QE drive the chosen robot after Force
Start, B kicks once per press, F the dribbler, stopped on key release, Halt or 0.3 s without a
message. The Robots panel shows each robot's share of its commands answered over the last 2 s,
red under find_robots' 80%. A demo switch drops the last demo's dashboard routes, so its
panels (Strategy, Teleop) never act on a closed runner.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@isaac0804 isaac0804 changed the title Field day 2026-10-27: real-robot restarts, calibration CLI, log findings Field days 2026-10-27..29: real-robot restarts, robot profiles, dashboard demos Oct 29, 2026
isaac0804 and others added 12 commits October 29, 2026 15:56
REAL_PARAMS back to 1 m/s, 4 rad/s, 50 rad/s^2; VISION_LOST_S and
CAMERA_SILENT_S 0.5 -> 0.1 s. Drops the test that pinned 0.2 s gaps as OK.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… body inside, brakes at 1.5

The fence capped speed towards the half-way line from the robot's last seen
position only: a robot already driving at the line kept its speed for the
~0.1 s before the cap acted. It now extrapolates by the measured velocity
over DELAY, stops the centre MARGIN (0.1 m) inside the line, and brakes at
1.5 m/s^2 (sideways calibration measured 1.6-1.9; it assumed 2.0).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t of the bottom-right quarter's lower half

StrategyRunner(keep_out=...) takes rectangles (vision coordinates). The field
fence caps each robot's speed into them as at the half-way line (measured
velocity, 0.1 m margin, corners included), in every mode; the field check fails,
so HALTs real play, for a robot inside; the Live tab draws them. KEEP_OUT in
examples/shooting_drill.py, passed by every example.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ball from vision

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed every calibration at once

Today's logs: 50-75 vision packets/s (92 on 10-27/28), a gap over 0.1 s every
second; each calibration HALTed 0.10-0.12 s after the last packet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on-loss limit 0.3 s

Calibration has no dashboard to press Force Start on, so one auto-HALT (its
first tick, before vision arrived; or a vision gap) held the robot still for
the rest of the run while the step's prompt repeated.
StrategyRunner.resume_after_auto_halt Force Starts once nothing fails; the
interrupted step starts over. VISION_LOST_S, CAMERA_SILENT_S 0.5 -> 0.3 s for
today's feed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…; calibration fences the field lines

Each auto-HALT restarted the interrupted move where the robot stood: a half-done
'back' drove on past its start, and three pauses took a robot 0.14 m past the
goal line. A pause within a there-and-back (or turn and return) now drives the
robot back to the pair's start, rests, and replays the pair. Calibration also
fences the touch and goal lines (StrategyRunner(fence_field_lines=True)).
The return half of each pair is named '... back'.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ld (DRILL_GOAL)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the dribbler's side turning at 0.7-0.95

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@isaac0804
isaac0804 enabled auto-merge (squash) November 2, 2026 23:40
@isaac0804
isaac0804 disabled auto-merge November 2, 2026 23:44
@isaac0804
isaac0804 merged commit cffe3a2 into main Nov 2, 2026
4 of 5 checks passed
@isaac0804
isaac0804 deleted the field-day-20261027 branch November 2, 2026 23:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

release:minor Minor changes to main

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant