What the first research missed — and what it got wrong
Round 1 gave us the V-seat, the release rollout, and the steal-sheet that became master75. Round 2 hunted the seven questions it left open. Three findings change the plan, three questions have no published answer at all, and one celebrated technique turns out to never have been built.
System: vision-only · no force sensor · 5-DOFCups: 9 cm tall · 7.5 cm rim · 5.5 cm baseSim = ground truthSources read 2026-11-12
Why this round mattered — look at F and H
Your screenshot from the live build, 12 Nov. Cup C is bang on. Cups F and H are off their wireframe targets — exactly the failure this research explains.
What you spotted: F (teal ring) and H (green ring) sit off-centre in their target wireframes while C is centred. In a V-seat, that sideways offset — not tilt — is the margin that matters, and it compounds upward: a base-row error becomes the seat error for the row above. The overhead inset labels are also piled on top of each other and unreadable — a separate UI bug from the placement bug.
Already applied in master75 (v4.24)
V-seat centreline targeting — aim at live supporter positions, not the nominal slot
Pre-release sim rollout — clone, release, step 2 s, re-seat once if it tips
Table-setdown fallback when the platform is blocked
From parent build log, 12 Nov — master75 check running at time of research.
What Round 2 says to do next (v4.25 candidates)
Extend the rollout to the whole trajectory — descend + release + retract, not just release (Finding #1)
Measure the in-gripper offset before placing — F/H offset starts at the grasp (Q5)
Don't build the slow-release probe as a primary check — it has no published parameters; the clone already answers it exactly (Q3)
Drop the 7 cm/s cap as a target — use non-uniform timing instead (Q7a)
The seven gap questions — verdicts at a glance
Q1V-seat thresholdNo published no.
Q2The descend itselfTechniques, no speed
Q3Slow-release probeNever built
Q4Full-trajectory verifyWell published ✓
Q5Grasp-offset measure≤2 mm published
Q6Recovery thresholdsNo thresholds
Q7Contradictions3 corrections
Ranked top 5 — new findings Round 1 didn't have
Ranked by expected impact for this system. Every figure below is as reported, with its caveat.
#
Finding
Why it matters here
Source
1
Full-trajectory clone verification is published practice — simulate descend + release + retract with checkpoint predicates; execute live only if the clone passes.
Catches descend-shove, which release-only rollout cannot see. In this sim it is exact, not approximate.
COBRA-PPM's 94.2% is sim-only, single-cube, vs a naive baseline — real demo was 4/5 trials on easier towers.
Stops us anchoring cup expectations on 94.2%. Transferable parts are its thresholds (τ = 0.8 / 0.40) and noise scale (σA = 1.57 cm).
Cannizzaro et al. 2024/25, arXiv:2403.14488 v3
4
Faster, non-uniform timing beats the 7 cm/s cap — 3-cup pyramid in 3.50 s at 29/30 trials.
Free transit can be fast; only the final approach needs to be slow. Uniform speed-up hurts; low jerk helps.
Moon et al. 2026, arXiv:2610.32676; RGB-Stacking task.py via DeepWiki
5
In-hand pose measurable to ≤2 mm / <1° with multi-view vision, 95% compensation.
Direct fix for the F/H problem: measure offset, correct the place target. Single wrist view alone is ~9 mm / ~20° — an order worse.
Chen & Li 2023, Manuf. Lett. 35:40–48; Nonnengießer & Peters 2025, arXiv:2506.10787
Honourable mentions: Im2Contact vision-only contact localisation (CoRL 2023); Saito et al. impact-force < one cup-weight criterion (2012); Jenga two-tier thresholds 0.32 N / 0.18 N (2023); GelSight <1° “stable” angle scale.
Interactive: the friction lab — Q1
No published lateral-offset threshold exists for a cup in a V-seat. The closest published number is a friction threshold. Drag μ and see where polystyrene sits.
Published PS–PS values span 0.2–0.5 depending on surface condition (Marks' Handbook Table 3.2.6; one ScienceDirect Topics table lists 0.20). At 0.20 the seat is below requirement — friction-marginal, not friction-safe. Failure modes from the analogue: μ₁ too low → top squeezes down; μ₂ too low → bases skid apart.
Our derivation — not published: With rim radius 3.75 cm, base 2.75 cm, taper ≈6.3°, supporters 7.5 cm apart, the seated cup rests on the rim lips, nestled by millimetres. Margin is anisotropic: centimetre-scale along the seat line, millimetre-scale perpendicular to it. That is why C can look fine and F/H — offset sideways — become the problem later. The true threshold depends on unmodelled rim-lip geometry, which is why sim measurement, not a formula, is the answer.
Sources: physicsforums.com/threads/coefficient-of-friction-between-cylinders.795163 · NPTEL Mechanics of Solids L12 (max supporter separation) · Jánoš et al. on support-polygon degeneration to a line · Fearing 1984, MIT AI Memo ADA148962
Finding #1 in detail — verify the whole trajectory (Q4)
Round 1 verified only the release. The literature verifies everything. Tap each stage to see its checkpoint predicate (CoEx 2026 term).
Approach
Side-approach, then purely vertical descend (MEAM 5200). No sweeping over the tower.
Descend ⚠
Checkpoint (a): no supporter moves > threshold (e.g. 1 mm) in the clone during descend.
Release
Checkpoint (b): placed cup stable after release in the clone.
Retract
Checkpoint (c): retract does not disturb the stack. Execute live only if all pass; otherwise re-plan.
Approach — published basis: Kunze & Beetz (2015) originated simulation-based projection: “By envisioning the outcome of actions before committing to them, a robot is able to… prevent itself from ending up in unwanted situations.” CoEnv/CoEx (2026): twin-verified trajectory 70% vs 50% without verification on five real multi-arm tasks.
Also: Robot-GST 2026 arXiv:2609.33872 · VPTwin 2026 arXiv:2609.33104 · Grounded-Dreamer 2026 arXiv:2610.13164 · RoboStereo TTPA 2026 arXiv:2603.12639 (learned world model — has physics-hallucination caveat our physics clone does not)
Q2 · The descend itself — the failure Round 1 never covered
Three rules to adopt
Side-approach, then vertical-only descend — UPenn MEAM 5200: “A fast landing with high impact could cause lower-level blocks to slide, rotate, oscillate…”
Peak contact force < one cup-weight — Saito et al. 2012, J. Robot. Mechatron. 24(1):28–36, DOI 10.20965/jrm.2012.p0028. Shape the deceleration profile; verify in the clone.
Stop at first contact, re-verify — Lerner quasi-static lowering until resistance; Im2Contact (Kim et al., CoRL 2023) for vision-only contact localisation from a single RGB-D view, even under occlusion.
What else the literature gives
Impedance probing — Ota et al. 2024, MERL TR2024-018: push −Z for T s / d mm at 10 Hz under impedance control; authors warn vision alone may be insufficient due to occlusion, and line-contact estimation is hard “where the tower is very unstable.”
Weight-compensating placement — FerRobotics US 10,449,675: pneumatic force ≈ object weight, so effective descent force ≈ 0.
Still no cup-specific descent speed / force / trajectory published
Three corrections to Round 1 (Q3 + Q7)
Correction 1 · Q3
The slow-release probe was never built
Lerner et al. (2024, arXiv:2404.17668) mention it in one sentence of Future Work: “One possibility is to slowly release the objects…” No opening amount, watch time, tip threshold, or on-fail action exists anywhere. Round 1 presented it as an established technique — it isn't.
Lerner's actual numbers: wait 0.1 s, average 0.5 s; press 10 N; shifts within 1 cm in one iteration for large offsets; only 17% success when small-offset adjustment is needed — exactly our V-seat regime.
Our sim proposal — NOT published: open ~0.5–1 mm per step, watch ~0.5 s/step, fail if move >1 mm or tilt >1° (GelSight scale, Takahashi et al. 2024, arXiv:2403.19129), re-seat on fail, abort to re-place after 2 fails. In our system the clone rollout already answers stability exactly, so a probe is a physical cross-check, not the primary test.
Correction 2 · Q7a
7 cm/s is not an optimum — it's a benchmark limit
0.07 m/s is RGB-Stacking's 4D Cartesian effector limit for all motion (task.py line 76; yaw 1.0 rad/s). Its real-robot setup also wasn't vision-only — it used a Robotiq FT 300 wrist F/T sensor (RoboCat, arXiv:2306.11706 §C.1), so the 10 cm retract rule isn't a vision-only result either.
Correction 3 · Q7b/c
CoM-in-polygon is necessary, not sufficient — and 94.2% doesn't transfer
Friction cones are an independent condition (Fearing 1984); supporters are free bodies, not fixed ground; dynamics aren't captured. COBRA-PPM: sim-only, K=1 single 7.5 cm cube, naive baseline, real demo 4/5 on easier towers, no descend modelling. Its 88.6% prediction accuracy (1,000 towers, precision 0.955, recall 0.868, AUC 0.961, τ=0.40) is the more transferable figure — and irrelevant here, because our clone gives ground truth.
Speed evidence — faster wins, if timing is non-uniform
Moon et al. 2026, Learned Speed Prior (arXiv:2610.32676), real Franka, 3-cup pyramid:
Retimed (learned prior)
3.50 s · 29/30
Default policy timing
6.02 s
SAIL (same task)
14.71 s · 0.40
Retimed: grasp motion 1.57× faster, transit 1.28×, up to 52% less jerk; 120 RoboLab sim tasks 53.3% vs 46.2%. Uniform acceleration performed worse.
Q5 · Grasp offset — the F/H fix, with numbers
A cup grasped off-centre lands off-centre. Scale check first: COBRA-PPM measured placement noise σA = 1.57 cm (Y up to 2.77 cm) — an order of magnitude larger than a millimetre V-seat margin.
V-seat margin (perp.)
~mm
Multi-view corrected error
≤2 mm
Single wrist-view error
~9 mm / ~20°
Uncorrected placement noise σA
1.57 cm
Worst axis (Y)
2.77 cm
Method that works
Chen & Li 2023 — multi-camera reconstruction of simple features (a cup's circular rims qualify): max remaining error ≤2 mm, <1°, 95% compensation. Use wrist camera from 2+ arm poses, or wrist + overhead, as the multi-view set.
Manufacturing Letters 35:40–48, DOI 10.1016/j.mfglet.2023.07.020
Reduce it at the source too
Pre-grasp alignment — leave a gap, proportional-control to the cup centroid before closing (Austin Villa @ Home 2020, UT Austin)
Fixturing — seat against a known surface / use gravity + environment geometry to drive uncertainty to sub-mm (>95% in sim + real, Connected Papers record)
Baseline caution: Nonnengießer & Peters 2025 visual-tactile fusion still averaged 7.5 mm / 16.7° under occlusion
Q6 · Recovery — nudge, re-place, or rebuild?
No published lean / error / row thresholds exist
Published rule shape (no numbers)
Distance-to-goal under a budget — Namasivayam & Paul 2024, arXiv:2405.18948: pick the recovery whose resulting state minimises distance-to-goal given a replanning budget. Motivating example is literally tower collapse.
Recova — Liu et al. 2026, arXiv:2610.01178: restorative / enabling / adaptive skills; real robots 23.8% → 77.5% → 87.5% with recovery skills; human intervention 87.5% → 0% over four rounds. No numeric thresholds — a VLM monitor decides.
Lerner backtracking — remove last rock, try a new sequence, backtrack further if exhausted. Trigger is “loop failed to converge,” not an angle.
Transferable template: Jenga two-tier thresholds
Marchionna et al. 2023, Sensors 23(2):752 — 0.32 N aggressive early / 0.18 N cautious late; push at 20 Hz, abort on high force; 80% of attempts successful. Template: be aggressive when little is built, cautious when rows are at stake.
Our starting thresholds — NOT published, calibrate in sim: nudge only if still seated and lean < ~5°; re-place if left the seat or lean ≥ ~5°; rebuild affected rows if a supporter shifted > ~2 mm or lean is in the base row. Never nudge a V-seated cup without clone-verifying the nudge first.
What remains genuinely unknown after both rounds
Six items. None can be answered from the literature — each must be measured in the sim.
1
Cup-specific V-seat offset & tilt thresholds. Depends on rim-lip geometry + friction. Sweep offset/tilt in the clone and record the topple boundary.
2
Vision-only slow-release probe parameters. No implementation exists to copy. Design + tune in sim (proposal in Q3).
3
Nudge vs re-place vs rebuild cutoffs. Clone-based distance-to-goal is the principled substitute; thresholds need calibration.
4
Cup-specific descent speed, release height, settle time. Round 1's gap stands. Nearest constraints: Saito force criterion + non-uniform timing.
5
Effective rim-lip contact geometry & friction behaviour. Determines wedge amplification; must come from the sim's contact model or measurement.
6
Can vision alone detect V-seat instability before release? MERL 2024 says maybe not (occlusion); Im2Contact 2023 suggests contact can be localised. Unresolved — the clone rollout sidesteps it here.