← Object gallery

Da Vinci Recursive Improvement CAD Harness

Parallel-jaw gripper · 8 GPT-6 Astra iterations

100 N pinch · 4 load cases · 3 parts
Loading CAD…
Iteration 07 · Stepped knee-braced mast

Self-improvement methods

The agent designs the finger’s rib layout, learns from measured loads and reuses its own tools. Model weights stay fixed.

1. Reflect and remember

Review stress, displacement and clearance results. Retrieve prior attempts with Atlas Vector Search and carry lessons forward.

2. Create and reuse tools

Write a member-sizing utility, pass six independent checks and execute it before subsequent designs.

3. Evaluate and revise

Change rib connectivity, dimensions and depth. Keep the lightest jaw pair that passes every fixed check.

Research evidence

Reflection and memory. Closing the Consistency Gap · 8 Sep 2026 ↗ reported +16 percentage points in AppWorld tasks succeeding on all five runs, using stored diagnostic guidelines with ReAct/GPT-4.1; +13 points on similar tasks.

Reusable skills. SkillAlchemy · 24 Aug 2026 ↗ reported +19.9 percentage points pass rate over execution without skills across 87 SkillsBench tasks, by creating reusable skill packages from source material.

Evaluation and revision. AIDE² · 22 Sep 2026 ↗ found 7 successive agent improvements in 8 days by testing changes to its own code. Gains transferred to four held-out benchmarks.

Recent preprints; results are from other tasks, not validation of this CAD harness or measurements of each method’s contribution here.

MongoDB Atlas

Documents store attempts, evaluations, tools and policies. Vector Search retrieves prior results. GridFS preserves CAD files and source snapshots.

Agent harness setup

Generate, evaluate, reflect

Astra uses high reasoning to choose nodes, rib connections and rectangular sections within fixed carriage and contact interfaces. CadQuery builds the guide base and two mirrored jaws in an isolated Docker container.

A separate evaluator verifies the exported STEP geometry, checks collisions at nine openings from 20 to 60 mm and tests contact with 20 mm and 60 mm sample gauges. A linear 3D beam-frame solver evaluates pinch, payload, lateral and combined loads.

A review-agent call interprets every measured result and saves the next working policy. After the first two attempts, a generated sizing utility passes independent tests before reuse. Atlas memories and actual tool outputs enter later design prompts. Git versions source and context.

Fixed acceptance checks

Nominal stress ≤ 80 MPa, tip displacement ≤ 0.25 mm, member buckling factor ≥ 2 and running clearance ≥ 0.25 mm. The objective is moving jaw-pair mass; the unchanged guide base is included only in total assembly mass.

Measurement scope

Mass is measured from STEP volume at nominal aluminium density. Structural results use ideal rigid beam joints and fixed carriage roots, with loads applied at the defined tip node. Pad offset couples, fillets, bearings and contact are not modeled. The actuator, friction pads, fatigue, local shear and torsional stresses are outside this study. The drone and mounting adapter are display context; installation and flight loads have not been evaluated.

This sequential study uses the same archive, memory and sandbox services as the full multi-agent workbench. The workbench’s Atlas triggers drive its two-specialist evaluation/reflection queue; this study uses a Python loop. Recorded API usage: $4.39 within a $25 cap.

All 8 design iterations

Moving jaw mass · fixed guide base excluded
BEST PASSING JAW PAIR523.6 → 193.2 g63.1% moving mass reduction

Same checks for every design
Displacement ≤ 0.25 mm
Stress ≤ 80 MPa

01

Conservative baseline frame

Baseline
Loading CAD…
Iteration 01 · Conservative baseline frame

9 ribs per jaw · 6 frame nodes · 24 mm extrusion

Moving mass
523.6 g
Total assembly
887.5 g
Displacement est.
0.002 mm
Stress est.
1.9 MPa

4 loads + 9 openings passed

Agent reflection

Use the exact supplied conservative baseline for initial four-load-case screening, preserving the fixed carriage, pad, and running clearances.

Baseline passes at 523.6 g moving mass. Combined loading controls stress (1.891 MPa in 0–4) and deflection (0.00242 mm); pinch controls buckling (factor 9542 in 1–5). These results support testing a substantially simpler topology.

Test one full-height braced bay: retain nodes 0–3 with edges 0–2, 1–3, 2–3, 0–3 at current sections. Preserve fixed interfaces; recheck all four loads, longer-member buckling, STEP mass and travel clearance before claiming improvement.

Initial topology explorationSTEP
02

Tapered triangular frame

Improved
Loading CAD…
Iteration 02 · Tapered triangular frame

5 ribs per jaw · 5 frame nodes · 10 mm extrusion

Moving mass
207.1 g
Total assembly
570.9 g
Displacement est.
0.047 mm
Stress est.
15.9 MPa

4 loads + 9 openings passed

Agent reflection

Replace the heavy ladder with a 10-mm-deep tapered triangle using 4-mm uprights, 5-mm diagonals and a 3-mm midpoint tie, preserving fixed interfaces pending four-load-case and clearance screening.

Triangle passes at 207.1 g moving mass versus 523.6 g baseline. Combined loading controls: 15.94 MPa in lower upright 0–3, 0.0470 mm tip deflection, and buckling factor 188.9 in lower diagonal 1–4. Clearance is 0.30 mm at all nine travel samples.

Test two unsplit root-to-tip legs: remove tie 3–4 and midpoint nodes; use edges 0–2 (4 mm) and 1–2 (5 mm), retaining 10 mm depth and fixed interfaces. Recheck all four loads, full-length buckling, lateral stiffness, STEP mass and travel clearance.

Initial topology explorationSTEP
03

Slim two-leg triangle

Improved
Loading CAD…
Iteration 03 · Slim two-leg triangle

2 ribs per jaw · 3 frame nodes · 10 mm extrusion

Moving mass
198.5 g
Total assembly
562.3 g
Displacement est.
0.062 mm
Stress est.
20.3 MPa

4 loads + 9 openings passed

Agent reflection

Use the prior 15.94-MPa combined-load stress margin to replace the tied, split frame with direct 3-mm upright and 4-mm diagonal ribs, preserving fixed interfaces pending four-load-case, full-length buckling and travel screening.

Direct legs pass at 198.45 g, saving only 8.60 g over the tied frame. Combined load controls upright 0–2 stress (20.28 MPa), deflection (0.0616 mm) and diagonal 1–2 buckling (24.08). Clearance stays 0.30 mm at nine samples.

Benchmark a Y graph: retain roots/tip and join both roots at (8,35), with one shared trunk to the tip. Size the trunk for combined bending, not the former upright’s axial role. Compare four loads, buckling, STEP mass and nine-position travel; do not assume savings.

Saved sizing tool usedSTEP
04

High-junction Y frame

Improved
Loading CAD…
Iteration 04 · High-junction Y frame

3 ribs per jaw · 4 frame nodes · 10 mm extrusion

Moving mass
193.5 g
Total assembly
557.4 g
Displacement est.
0.120 mm
Stress est.
54.4 MPa

4 loads + 9 openings passed

Agent reflection

Target lower mass than the 198.45-g passing frame with 3-mm root ribs merging at (8,60) into a 4-mm shared trunk, limiting its new bending span to 10 mm while retaining fixed interfaces pending four-load stress, deflection, buckling and STEP/travel screening.

Y passes screening at 193.52 g, saving 4.93 g but nearly doubling deflection versus prior direct legs. Combined load controls root 0–3 stress (54.39 MPa), tip deflection (0.1196 mm) and diagonal 1–3 buckling (factor 12.16).

Benchmark two direct 3-mm root-to-tip legs at 10-mm depth, removing the junction and shared bending span. Compare mass–stiffness tradeoffs without assuming savings; preserve interfaces and recheck four loads, full-length buckling and STEP/travel.

Saved sizing tool usedSTEP
05

Minimum-section direct legs

Passed
Loading CAD…
Iteration 05 · Minimum-section direct legs

2 ribs per jaw · 3 frame nodes · 10 mm extrusion

Moving mass
194.6 g
Total assembly
558.4 g
Displacement est.
0.071 mm
Stress est.
22.1 MPa

4 loads + 9 openings passed

Agent reflection

Replace the Y junction and shared bending span with direct 3-mm root-to-tip ribs to address the prior 54.39-MPa combined-load root stress and benchmark mass against the 193.52-g best design, preserving fixed interfaces pending four-load stress, deflection, buckling and STEP/trave

Direct 3-mm legs pass at 194.56 g, 1.04 g above the Y, but reduce combined-load stress to 22.15 MPa (0–2) and deflection to 0.0714 mm. Combined compression in diagonal 1–2 controls buckling at 10.12; sampled travel clearance remains 0.30 mm.

Test a 6-mm-deep tied triangle: split both 3-mm legs at mid-height and add a 3-mm midpoint tie. Seek mass savings without a shared bending trunk. Preserve interfaces; recheck lateral/combined deflection, effective buckling lengths and STEP/travel.

Saved sizing tool usedSTEP
06

Canted minimum-width Y

Improved
Loading CAD…
Iteration 06 · Canted minimum-width Y

3 ribs per jaw · 4 frame nodes · 10 mm extrusion

Moving mass
193.4 g
Total assembly
557.3 g
Displacement est.
0.143 mm
Stress est.
74.8 MPa

4 loads + 9 openings passed

Agent reflection

Replace the direct legs with an all-3-mm Y merging at (10,59), retaining the contract-required 10-mm depth rather than 6 mm and limiting the shared bending rise to 11 mm to target the 193.52-g best while preserving fixed interfaces, pending four-load and travel screening.

Passes at 193.38 g, just 0.14 g below the prior Y. Combined loading controls trunk 3–2 stress (74.77 MPa) and deflection (0.1427 mm); pinch controls diagonal 1–3 buckling (12.89). The tiny saving comes with substantially higher bending stress.

Use the proven direct 3-mm legs as the next baseline: 1.17 g more mass halves deflection and lowers peak stress to 22.15 MPa. Avoid further Y-junction nudges; retain 10-mm depth and require a meaningful load-path change before repeating four-load and CAD travel screening.

Saved sizing tool usedSTEP
07

Stepped knee-braced mast

Best
Loading CAD…
Iteration 07 · Stepped knee-braced mast

5 ribs per jaw · 6 frame nodes · 10 mm extrusion

Moving mass
193.2 g
Total assembly
557.1 g
Displacement est.
0.246 mm
Stress est.
77.1 MPa

4 loads + 9 openings passed

Agent reflection

Replace the Y’s 74.77-MPa combined-load trunk with a lower knee brace and selectively reinforced mast, targeting lower mass than 193.38 g while preserving fixed interfaces, pending four-load stress, deflection, buckling and STEP/travel screening rather than certification.

Screening passes at 193.19 g, only 0.20 g below the Y. Combined loading controls lower mast 0–3 stress (77.11 MPa), deflection (0.2463 mm) and diagonal 1–4 buckling (14.97). Lowering the brace worsened bending despite selective reinforcement.

Restore two direct 3-mm root-to-tip legs at 10-mm depth, eliminating the shared cantilever. Prior screening gave 22.15 MPa and 0.0714 mm for just 1.37 g more. Preserve interfaces; recheck four loads, full-length buckling and STEP/travel rather than trim another Y.

Saved sizing tool usedSTEP
08

Inclined Shared Spine

Passed
Loading CAD…
Iteration 08 · Inclined Shared Spine

3 ribs per jaw · 4 frame nodes · 10 mm extrusion

Moving mass
193.3 g
Total assembly
557.1 g
Displacement est.
0.246 mm
Stress est.
79.7 MPa

4 loads + 9 openings passed

Agent reflection

Replace the stepped mast, whose combined-load stress reached 77.11 MPa and deflection 0.2463 mm, with two inclined 3-mm root branches feeding a 4-mm shared spine, targeting lower mass than 193.19 g while preserving fixed interfaces pending four-load, buckling and STEP/travel scre

Screening passes but misses the mass target: 193.27 g, up 0.08 g. Combined stress in root 0–3 reaches 79.71/80 MPa. Pinch controls deflection (0.24634/0.25 mm) and diagonal 1–3 buckling (15.73). The inclined Y leaves little stress or stiffness margin.

Remove the shared cantilever: restore direct 3-mm edges 0–2 and 1–2 at 10-mm depth. Prior screening gave 194.56 g, 22.15 MPa and 0.0714 mm—a stronger tradeoff than junction nudges. Preserve interfaces; recheck four loads, full-length buckling and STEP/travel.

Saved sizing tool usedSTEP
Earlier sensor-mount study ↗Full harness ↗