Continual Learning — Progress Dashboard

iter_001 — fail

Failure analysis

# iter 001 — observation run (no failure)
Pure sensing: printed EE pose + coarse deprojected grids from all three cameras.
Learned: base cam faces the microwave head-on; front face is a plane at x≈0.70;
image-left = +y; EE home (0.577, 0.001, 0.83), quat (.5,.5,.5,.5); floor z≈0.01.

Proposed solution

Focused scan of the handle region and a horizontal slice of the door to get
handle position, protrusion, and door extents before attempting a pull.

iter_002 — fail

Failure analysis

# iter 002 — observation run (no failure)
Mapped geometry: door plane x≈0.702; handle bar at x≈0.65-0.665 (protrudes
~4.5 cm), y≈-0.205, z 0.52-0.78; control panel y<-0.25; cabinet spans y ±0.37.
Assumed hinge at left door edge y≈+0.35 (WRONG — later fitted to y≈0.11).

Proposed solution

Calibrate fingertip offset by pressing the door window (move_to residual as
contact signal), grasp the handle bar, pull along a modeled hinge arc with the
wrist yaw synced to door angle.

iter_003 — fail

Failure analysis

# iter 003 — grasp missed (progress 0.0)
Probe worked: fingertip offset L=0.051 m (closed fingers touch door at
ee_x=0.651). Grasp closed with tips only ~1 cm past bar center; the wedge tips
squeezed the bar out ("watermelon seed"). The empty blade then swept the arc;
at theta>=0.3 the commanded poses also stalled (model arc mismatched reality).

Proposed solution

Approach 1.1 cm deeper so tips are ~2.5 cm past bar center; add a tug test to
verify the grasp before pulling; retry deeper if it slipped.

iter_004 — fail

Failure analysis

# iter 004 — grip torn off mid-pull (progress 0.24)
Deeper grasp held (tug confirmed). Door opened but the commanded arc (hinge
guess y=0.35, r=0.557) had ~1.75x too large a radius: the door swung much
faster than the commanded angle, the wrist desynced, the gripper got dragged
down the bar and popped off; door settled at 0.24 rad. KEY: the door holds its
angle after release (hinge friction, no spring-back). Circle-fitting the
attached-pull EE track recovered the true hinge: (0.650, 0.110), r=0.318.

Proposed solution

Re-run the pull using the fitted hinge (0.650,0.110) r 0.318; arc reaches only
comfortable workspace (x>=0.35) so pull to theta ~1.2 in one go.

iter_005 — fail

Failure analysis

# iter 005 — gradual slip (progress 0.53)
Fitted-hinge arc tracked much better, but residual model error still made the
bar slide slowly through the fingers (tracking err grew 0.014->0.036 from
theta 0.8-1.16, then the bar popped out; free arm then tracked exactly).
Open-loop precision is not sufficient; need mid-task re-measurement.

Proposed solution

Two-stage: pull to ~0.5, release, park arm, re-scan with base cam (fit the
open door line -> true angle; find handle bar), recompute hinge in closed
form, re-grasp at the measured bar pose, pull the corrected arc.

iter_006 — fail

Failure analysis

# iter 006 — bad bar detection -> arm crashed through door (progress 0.34)
Line fit worked (phi=0.36 measured), but the "bar cluster" detection averaged
scattered points (door edge side-face + parked-arm points) into a garbage
position; the re-grasp move_to then swept the arm through the open door and it
stayed wedged/flailing for the rest of the episode.

Proposed solution

Iterative servo loop: short pull segments, rescan between each; grip point
derived from the door line free edge; collision-safe routing (orientation
changes far from the door, staged approach); abort segment on error spikes.

iter_007 — fail

Failure analysis

# iter 007 — edge-derived grips missed the bar; empty sweeps CLOSED the door
(progress 0.03). The handle occludes the door's free edge from the base cam,
so edge-98th-percentile grip points landed ~4 cm hinge-ward of the bar. Every
re-grasp closed on air (no grip verification!), and sweeping the closed empty
blade along the arc nudged the door progressively shut (0.28->0.22->0.12).

Proposed solution

Detect the bar DIRECTLY as a tight cluster 3-6.5 cm in front of the fitted
door line (median-gated); park the arm lower/further out of the scan box;
verify every grasp with a radial tug (held bar cannot move radially); bias the
arc radius inward 1.2 cm to press the bar into the palm; skip pull on empty
grasp.

iter_008 — success

Failure analysis

# iter 008 — SUCCESS (progress 1.50, door fully open)
Servo loop: closed grasp + pull (door ~0.38); round-2 grasp was empty but the
radial tug caught it (no destructive sweep); rescan phi=0.19, re-grasp, pull
-> 0.755; re-grasp (model grip; bar occluded at large angles), pull -> door
hit its fully-open stop at 1.50 rad. Final scans confused by the fully-open
door (line fit garbage / 0 points) but the loop exited safely. Success checker
requires >=0.8 rad after release: ended at 1.50.

Proposed solution

Solved. Distilled into skills/ (hinged-door pull toolkit, depth line/cluster
scanning, grasp verification) and world_memory/ (scene geometry, embodiment
constants, articulated-object handling lessons).

iter_001 — fail

Failure analysis

# b iter 001 — observation. Face plane x~0.415, cabinet y ±0.227, z 0.40-0.66.
No protruding handle visible head-on; full-width base strip at z 0.39-0.40
(x 0.393). Door window recessed 4 mm (y -0.11..+0.19, z 0.44..0.60).

Proposed solution

Fine relief profiling (1 cm binned x(y), x(z)) to find lips/ridges/rods.

iter_002 — fail

Failure analysis

# b iter 002 — relief profiles: face is essentially FLUSH (<=4 mm relief).
Only the bottom strip protrudes (2.2 cm). No graspable handle found by depth.

Proposed solution

Physical exploration: blade sweeps (depth underestimates protrusions),
strip/table probes, pinch-and-pull tests.

iter_003 — fail

Failure analysis

# b iter 003 — sweeps found panel BUTTONS protrude ~1cm; several probes
crashed the arm (rolled-90 poses + low-z laterals). An accidental knock
opened the door 0.04 rad -> NO LATCH. progress 0.039.

Proposed solution

Controlled crack tests + plane-tilt scan (x=a+b*y+c*z) to identify the hinge
axis (side vs bottom).

iter_004 — fail

Failure analysis

# b iter 004 — every contact move crashed (y>0 cross-body with HOME quat AND
z<0.42 near the table front edge are unreachable/unstable zones). No tilt
detected; tests never touched the door. progress 0.

Proposed solution

Stay in safe pose space (y<=-0.05, z>=0.45); pitched-down blade for low
targets; strip drag to crack, then edge-hook arcs.

iter_005 — fail

Failure analysis

# b iter 005 — pitched strip approach still flailed, but the arm banged the
door to 0.198 (measured); hooks failed (routing through x~0.2 cross-body
folded configs). Door held 0.125 at end. KEY: video shows the LEFT (+y) edge
protruding -> door is RIGHT-hinged (hinge y~-0.11, free edge +0.19, r 0.30).

Proposed solution

Probe the suspected free-edge grip; hook it and drag along the +z rotation
arc with yaw following; measure between rounds.

iter_006 — fail

Failure analysis

# b iter 006 — JACKPOT discovery: probe/hook contacts at tips x~0.34-0.36 at
y~0.21 reveal a handle on the door's left SIDE face, OUTSIDE the base cam
FOV (cam only covers |y|<~0.21 at this range) — invisible to all depth scans.
Hook-drag from closed reliably opened to 0.29-0.34 (twice). But continuing
from phi~0.3 fails: the handle arc runs into the unreachable cross-body zone
(x<0.2, y>0.2) and the flailing arm knocked the door shut each time.

Proposed solution

Keep the proven hook-crack; then switch mechanism: pitch-1.4 blade from above
pressing the door INNER face at small radius (r 0.13) — that contact point
stays in the comfortable workspace through phi 1.2 (D-scene technique).

iter_007 — success

Failure analysis

# b iter 007 — SUCCESS (progress 0.896). The hook-crack drag alone carried
the door to 0.89 rad this run (cleaner initial routing gave the blade a
better bite on the side handle); measured phi 0.890 >= 0.85 so the wedge
press was not needed. Door held through release/retreat/settle.

Proposed solution

Solved. Distilled: side-handle hook-crack + inner-face wedge press reserve.

iter_001 — fail

Failure analysis

# c iter 001 — observation. Retro microwave: face x~0.4335, cabinet y ±0.25,
teardrop window door, 2 knobs right panel. Coarse scan found only a
full-width strip at z 0.39 (base).

Proposed solution

Relief profiling to find the handle.

iter_002 — fail

Failure analysis

# c iter 002 — profiles: thin vertical rod-like proud region near y -0.12
(x dips to 0.416-0.427), tilted (y -0.10 at z 0.47 to -0.15 at z 0.62);
also proud spots at y +0.11..+0.22 (window molding). All ~1-1.5 cm per depth.

Proposed solution

Task-A style servo loop grasping the rod at mid-height.

iter_003 — fail

Failure analysis

# c iter 003 — rod pinch at z 0.55 blocked ~1.5 cm short (rod crosses the
proud trim there); all rounds identical failures, tug caught every empty
grasp (no destructive sweeps). progress 0.

Proposed solution

Grasp low (z 0.47) where rod is clear of trim; add touch-probe to find the
rod x physically.

iter_004 — fail

Failure analysis

# c iter 004 — the direct probe-start move flipped the arm (x -0.045!);
staged approaches reached perfectly (err 0.004) but the pinch closed on
nothing: wedge-shaped tips cannot capture a thin shallow rod against the
face plane (squeezed out). progress 0.

Proposed solution

Physical probes (lateral sweep + x-probe) to get true rod geometry; test
right-hinge hypothesis (rod could be a hinge pin).

iter_005 — fail

Failure analysis

# c iter 005 — probes rewrote the map: the rod protrudes <1.2 cm (unusable),
but the thick rolled window MOLDING protrudes to x~0.396 (3.7 cm!) at
(y -0.06, z 0.467) — depth had underestimated ~2x. A cross-body move to
y +0.205 crashed the arm (workspace lesson). progress 0.

Proposed solution

Pinch the roll rolled-90 where it runs horizontal; +z tug (universal check);
pull left-hinge arc; decide hinge from measured motion.

iter_006 — success

Failure analysis

# c iter 006 — SUCCESS (progress 0.872). Rolled-90 pinch of the roll HELD;
left-hinge pull cracked the door (arc radius imperfect -> growing errors but
real opening; mid-run line-fit scan misleading because the fixed panel
dominates). The attempt-B probe then contacted the swung-out door edge at
x 0.30 and subsequent moves pushed it to 0.87 where friction held it.

Proposed solution

Solved. Lesson: pinch thick rolled moldings (rolled-90 for horizontal runs);
physical probes over depth estimates; scan boxes must isolate the door.

iter_001 — fail

Failure analysis

# d iter 001 — observation. Toaster-oven layout: flat face x=0.420 (y ±0.25,
z 0.38-0.76), horizontal bar handle x~0.383 z 0.66-0.68 spanning y ±0.19,
knobs on the top strip (z 0.72-0.75, x 0.411). Door = dark slab below the
bar -> drop-down (bottom-hinged) door suspected.

Proposed solution

Relief profiles to confirm; then bottom-hinge pull plan.

iter_002 — fail

Failure analysis

# d iter 002 — profiles confirm: perfectly flat face, bar at z 0.66-0.68
proud 4 cm with a clear slot behind, base strip z 0.39-0.40. Hinge assumed at
door bottom (z~0.41), axis along y.

Proposed solution

Pinch the bar rolled-90, pitch-synced arc pull; gravity should assist once
tilted (COM starts above the hinge).

iter_003 — fail

Failure analysis

# d iter 003 — rolled-90 pinch pose unreachable (controller settled 10 cm
off); scan-based angle fit worked (phi 0). progress 0.

Proposed solution

No pinch needed for a drop-down door: hook the bar from above with a
pitched-down closed blade and drag.

iter_004 — fail

Failure analysis

# d iter 004 — descent from above stalled at tip z~0.78: the box TOP SLAB
OVERHANGS the front face (invisible to the head-on cam); the blade landed on
it. Push fallback assumed phi>=0.1 and swept free air (door truly at 0).
progress 0.

Proposed solution

Press-rake: the bar is 3.7 cm FORWARD of the hinge axis, so pressing straight
down on it tips the door open. Approach from front-below the overhang.

iter_005 — fail

Failure analysis

# d iter 005 — press missed: the real bar front is ~0.372-0.375 and extends
z 0.65-0.685 (depth cluster was offset); the press descended millimeters in
front of it and the rake swept air. progress 0.

Proposed solution

Slide in through the corridor (below overhang, above bar), press down ON the
bar with contact verification, rake centered on the verified contact.

iter_006 — fail

Failure analysis

# d iter 006 — contact-verified press + rake opened to 0.45 rad. Gravity did
NOT finish (friction hinge holds any angle). Second rake missed (hinge model
error grows with angle; bar is a 2 cm target). progress 0.45.

Proposed solution

Servo rounds: refit hinge z from scans, re-acquire the bar by pressing to
contact before each rake span.

iter_007 — fail

Failure analysis

# d iter 007 — REGRESSION to closed: blind bar re-acquisition landed on the
door face, produced garbage deltas, and the correction-shifted rakes pushed
the door SHUT. Bar too small a target at unknown angles. progress 0.

Proposed solution

Press-slide the exposed INNER FACE instead (huge target; downward press gives
opening torque ~ s*sin(phi)); choose press radius from the scanned strip.

iter_008 — success

Failure analysis

# d iter 008 — SUCCESS (progress 1.571, fully flat). Bar rake to 0.47, then
inner-face press-slides: round 2 pushed past ~1.0 rad where gravity slammed
it fully open. Post-flat scans read garbage (door left the scan box) and
triggered harmless re-rakes. Door held at pi/2.

Proposed solution

Solved. Distilled: press-rake + inner-face press for bottom-hinged doors;
add flat-door detection to scans in future.

Human demonstration

iter_001 — fail

Failure analysis

# iter_001 — progress 0.37 rad (fail)
Demo-mimic hook-and-arc-pull. The exact demo EE grasp (0.652,-0.232,0.585) hooked
the handle and the arc around an assumed hinge (0.66,0.21) r=0.44 tracked to
theta~0.72, opening the door to ~0.37 rad. Then the fingers slid DOWN the vertical
handle bar (EE z 0.58->0.44) and off; the arm flailed toward unreachable arc
targets. The "push" phase then pressed the door's OUTER face, which pushes a door
CLOSED (outer-face normal points in the opening direction) — lucky it didn't
lose more. Compared to the demo: demo also releases the handle around ~0.4 rad
and finishes by pushing from the inner side.

Proposed solution

Split into two phases like the demo: (1) hook-pull only in the reliable range,
(2) get the gripper INSIDE the wedge between the ajar door and the cabinet and
push the INNER face around the hinge. Pushing inside the wedge is robust to
hinge-estimate error; pulling is not.

iter_002 — fail

Failure analysis

# iter_002 — progress 0.11 rad (fail)
Crack-then-push-sweep. The pull cracked the door (~0.35 rad) but the wedge entry
path, computed from a guessed geometry (plane x=0.66), hit the door's outer face
and pushed it nearly shut; the sweep then tracked free air (err~0.001-0.005 =
no contact). True front plane turned out to be x=0.70 (measured next run).

Proposed solution

Stop guessing geometry. Measure with the camera: deproject a depth grid to a
world point cloud, estimate the front plane (modal x), find the handle, and
after cracking the door fit the ajar panel to recover the hinge.

iter_003 — fail

Traceback (most recent call last):
  File "/home/r2d2/cl_demo/_harness/child_runner.py", line 38, in main
    program.solve(robot)
  File "/home/r2d2/cl_demo/workspace/open_microwave_v003.py", line 93, in solve
    p2 = scene_stats(pts2, "ajar")
  File "/home/r2d2/cl_demo/workspace/open_microwave_v003.py", line 34, in scene_stats
    box = (pts[:, 0] > 0.35) & (np.abs(pts[:, 1]) < 0.55) & \
IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed

Failure analysis

# iter_003 — progress 0.0 (fail, program error)
Vision worked: front plane X0=0.700, handle bar at (0.659,-0.21,0.59) (4cm proud
of the plane). But two bugs: (1) "improving" the grasp y to the measured bar
centroid (-0.21 vs demo -0.232) made the hook MISS entirely — the demo-exact EE
pose matters at the centimeter level; (2) the second point-cloud capture returned
an empty array and crashed the fit (unguarded reshape).

Proposed solution

Revert to the demo-exact grasp pose. Guard all cloud math. Use measured plane
x=0.70 to correct the pull-arc hinge.

iter_004 — fail

Failure analysis

# iter_004 — progress 0.0 (fail)
Approaching the pre-grasp point directly from the folded VIEW pose left the arm
in a bad configuration: move_to stalled 0.56m from the handle, no grasp, no
door opening. Routing matters: the working grasp approach starts from the HOME
region. The ajar "measurement" then fit garbage (it was seeing the handle bar
of a closed door); sanity gates correctly rejected it.

Proposed solution

Route VIEW -> HOME -> pre-grasp -> grasp. Mask the door-panel band at
x < X0-0.075 so the closed door's handle bar can't masquerade as an ajar panel.

iter_005 — fail

Failure analysis

# iter_005 — progress 0.06 (fail)
Grasp+pull worked (corrected hinge x=0.695): tracked to theta 0.7, door ~0.5.
But the fingers still slid off (z dropping past theta 0.8), and the ajar-door
measurement was polluted by 1100+ ARM points (the VIEW pose sits inside the
door z-band), skewing the hinge fit; the wedge entry collided and the sweep
pressed the outer face, re-closing the door.

Proposed solution

Stop the pull at theta 0.65 (before slip). Clean the measurement by diffing
clouds captured from the identical arm pose (only moved points = door).
Enter the wedge radially through its mouth, parallel to the door plane.

iter_006 — fail

Failure analysis

# iter_006 — progress 0.18 (fail)
Pull clean (door ~0.5, visible in video). Diff-based measurement still off
(hinge y 0.286, theta 0.61 overestimated): the radial mouth insertion (7cm
margin) clipped the door edge, deflected low, and the sweep dragged the OUTER
face, closing the door to 0.18. Root issue: fit bias from the handle points +
insertion margins smaller than the fit error.

Proposed solution

Enter the wedge where the margin is largest. Candidate: from above — the wedge
is open at the top. Also constrain the hinge fit to lie on the measured front
plane (1-parameter search) instead of extrapolating a free line fit.

iter_007 — fail

Failure analysis

# iter_007 — progress 0.006 (fail)
Self-inflicted: changed transit tolerances to 0.01; the VIEW pose is only
reachable to ~3cm, so both VIEW moves timed out in odd configurations, every
subsequent move started ill-conditioned, the arm flailed across the scene
(one swing shut the door), and the overhead approach overshot into the cabinet
top edge. Lesson: tol 0.01 only for the grasp; 0.03+ for transits.

Proposed solution

Restore tol 0.03. Drop cloud diffing entirely: choose a door-band mask
(x in [0.40, X0-0.06]) that geometrically excludes the arm, the closed door
plane, and the handle bar — no pose repeatability needed.

iter_008 — fail

Failure analysis

# iter_008 — progress 0.002 (fail)
Two compounding: (1) door-band z>0.44 swallowed the stand tabletop -> fit hit
the grid boundary -> fallback geometry; (2) clamping the overhead entry x
silently increased its polar angle, parking the descent OVER the door panel:
the gripper landed on the door top, slid off outside, and dragged the door
shut. Also: a 10cm descent step with a short timeout overshoots ~8cm.

Proposed solution

z-band 0.50-0.78 for the fit. Tactile descent (small steps, detect contact,
back off toward the cabinet on door-top contact). Slip-guard the pull.

iter_009 — fail

Failure analysis

# iter_009 — progress 0.0 (fail) — but the KEY measurement
The clean fit (spread 0.017 rad, 540 pts) revealed the TRUE geometry:
hinge=(0.70,0.35), edge radius ~0.58 — not (0.70,0.235)/0.47. This explains
every slip: arcs around the wrong hinge run ~2cm inside the true handle circle
by theta~0.7, jamming the fingers down/off the bar. In this run the descent
probe overshot, knocked the door shut, the sweep swept air, and near
(0.32,0.15) the arm hit an instability and flailed.

Proposed solution

Pull along the TRUE arc (hinge 0.70,0.35, r 0.584) — expect a deeper, cleaner
pull. Keep the EE out of the unstable near-base region (x<0.33, y~0.15-0.25);
success only needs 0.8 rad, so stop the sweep at ~1.1.

iter_010 — fail

Failure analysis

# iter_010 — progress 0.001 (fail)
True-arc pull: errors <=0.018 all the way to theta 0.62 — door ~0.55 open,
slip solved. But the overhead transit at z=0.90 crossed the open door's swing
sector: the door is TALLER than assumed (top ~0.85), the hand clipped it,
jammed, and the recovery sweep shut the door from outside.

Proposed solution

Never fly over the door. With the door ~0.5 open, enter HORIZONTALLY through
the wedge mouth at door-mid height (z=0.58) from the front-right; the path can
provably stay clear of the door line, the cabinet plane, and the handle.

iter_011 — fail

Failure analysis

# iter_011 — progress 0.002 (fail, one waypoint from success)
Everything executed perfectly (entry err 0.006, sweep tracked to theta 1.08)
— but in free air. The transit VIEW -> staging (0.45,-0.35) crossed the door
plane only ~4cm beyond the measured free-edge radius; the hand clipped the
open door en route and slammed it shut before entry.

Proposed solution

Route around the door's swept circle: VIEW -> (0.25,-0.30) -> (0.45,-0.35)
moves the door-plane crossing to radius ~0.78 (17cm clearance). Re-measure at
the staging point and re-pull once if the door reads closed.

iter_012 — success

Failure analysis

# iter_012 — SUCCESS: door 1.24 rad, stays open (needed >=0.8)
Final pipeline:
 1. Hook the handle at the demo-exact EE pose (0.652,-0.232,0.585); pull along
    the true handle arc (hinge (0.70,0.37), r 0.60) to theta 0.62 -> door ~0.44.
 2. Measure from VIEW: hinge-constrained fit gives hinge_y=0.380, theta=0.435,
    r_edge=0.606 (spread 0.031).
 3. Safe route around the door's swept circle to staging (0.45,-0.35,0.58);
    re-check door angle from there (still 0.435 — routing fix worked).
 4. Horizontal wedge entry at z=0.58: polar(0.45,0.25) then polar(0.38,0.25).
 5. Push-sweep the inner face at r=0.38 to theta 1.10; contact signature
    visible as rising move errors (0.017->0.092) after theta 0.94.
 6. Retreat radially + back through the wedge mouth at z=0.58.
Door settled at 1.2446 rad after release and retreat. Wall 6.5s, sim 31.6s.
Trajectory similarity to demo 0.877 (two-phase strategy mirrors the demo).

Proposed solution

Solved. Distilled reusable strategy into skills/ (hinged-door recipe, vision
fitting, motion gotchas) and scene/embodiment facts into world_memory/.

iter_001 — fail

iter_002 — fail

iter_003 — fail

iter_004 — fail

iter_005 — fail

iter_006 — fail

iter_007 — fail

iter_008 — fail

iter_009 — fail

iter_010 — fail

iter_011 — fail

iter_012 — fail

iter_013 — fail

iter_014 — fail

iter_015 — fail

iter_016 — fail

iter_017 — fail

iter_018 — fail

iter_019 — fail

iter_020 — fail

iter_021 — fail

iter_022 — fail