Research comparison • 18 Nov 2026Index evidence • Not live-verified
Simulation-first • World models • Kit-bashed hardware
Build the nervous system in sim first. The evidence says yes — but not losslessly.
You asked for 100 independent case studies on building a robot's perception, world model, control and planning in simulation before building the body — plus low-cost kit-bashed hardware. This page is that corpus, honestly audited, and put head-to-head against our own work: the browser physics sim, the pyramid stacking runs (master50 → master83), and the plan to kit-bash a cheap physical rig.
Corpus 100 entriesSplit 50 / 40 / 10 — on-topic / related / kaleidoscopeTags [published] [reported] [estimated] [unknown]Our baseline master83: 6/10 — best real score yet
The honest funnel — read this first
The brief said “100 independent.” The audit says that label is wrong — and that matters for how much weight to give any single claim.
Entries collected
100
Distinct contributions
98
Independent evidence units
73
Research programs
~48
Duplicates collapsed: SO-101 described 3× (R16–R18). Largest clusters tied at 7% each — SO-101 / LeRobot (7 entries) and ETH legged locomotion + Isaac Gym (7 entries). Top 5 clusters = ~30% of the corpus. Full audit in notes/04-audit.md. Two entries (OT17, OT44, FREEDERIA — 85% and 92% claims) are from an unverified outlet and are excluded from strongest findings.
Source note on the count: the audit file’s header first writes “97 distinct contributions,” then Step 1 recounts and restates it as 98 (“only the SO-101 triple collapses, −2”) — the figure report.md also uses. This page follows the revised final count (98); the 73 units and ~48 programs are unchanged in both versions.
Bottom line for our build
Randomization breadth beats simulator fidelity. Every success pairs sim with a transfer mechanism. Pure sim with no randomization fails.
DR-trained policies: 87–100% across randomization levels vs 40% nominal-trained (OT15)
Sim success is an unreliable predictor of real success (OT16)
Browser physics has zero validating primary sources in the corpus
The comparison you asked for
100 studies vs. our research
Our research = the browser-first stack you’ve been building: physics sim as the nervous-system testbed, trajectory / wrench / stability checks for stacking, creep and repair logic, and a low-cost kit-bash path to hardware. Here’s where the published corpus lands on each bet.
Our bet / finding
What the corpus says
Verdict
Build control & planning in sim before the body exists
Field default for Dactyl, ANYmal, Dreamer, modern VLAs. Massey digital-twin thesis (n=4, [published]): −1 month / NZD 8,080; −44% cost / NZD 9,380; −20.84% / NZD 42,500; −2 months / NZD 19,280. Virtual commissioning: 14 → 6 days on-site, 47 → 8 change orders, one $180k error caught (OT27, [reported]).
Confirmed
World model as the robot’s internal model — predict, then act
Ha & Schmidhuber V-M-C: CarRacing 906±21, VizDoom 1092±556 trained purely in dream [published] (OT10). DreamerV3: 150+ tasks, one config, first Minecraft diamonds from scratch (OT09). TD-MPC2 scales to 80–100+ tasks decoder-free (OT12). Cosmos Policy: LIBERO 98.5%, real ALOHA 93.6%, best-of-N +12.5 pts [published] (OT13).
Confirmed — architecture matters
Browser sim (Three.js / cannon-style) is enough to develop the nervous system
No retrieved primary source validates browser physics for robot control. Corpus validates MuJoCo / PyBullet / Isaac-class engines. Closest analogues: Webots (open, scriptable) and PyBullet (pip-install, ~30 s) for prototyping, MuJoCo as fidelity step-up. Report §4 & “Could not verify.”
Gap — treat as logic testbed
Trajectory check should veto a bad placement — master83: B caught a real 72 mm shove on supporters, build went ahead anyway, J + C fell; C’s repair never fired
Corpus agrees checks must gate action: OT30 zero-net-wrench release criterion — 100% basic stacking, only 17% on adjustment cases, 46 trials [published]. OT29 infers placement stability from tactile + F/T contact patch. OT31 uses symbolic thresholds (fn ≥ fmin; slip |v| ≥ vth → re-grasp) with [unknown] measured values. No plug-in threshold exists — you must measure yours.
Confirmed — and explains master83
Repair / recovery logic after a disturbance (creep test passed clean on master83; J carried back into slot)
Supported in principle: ANYmal fall recovery learned in sim (OT04), ADR widening as policy succeeds (OT03), DayDreamer learning on real robots within hours (OT11). But Dactyl shows cost: median 11.5 real rotations vs 50 in sim [published] — recovery that looks solved in sim overstates real by ~4× on contact-rich tasks.
Direction right, discount sim score
Sim score predicts real score — chase the sim number
Directly refuted. Abekenov SAC push (OT16): sim 79% sparse/dense, 67% with DR; real 80% straight-line / 50% overall for DR policy. Lowest-sim-success policy was the only deployable one. Lateral/diagonal pushes failed — outside training distribution.
Refuted
Kit-bash a capable arm for $100–500
Proven at component level. SO-101 pair $259 kit / $229.88 parts, 500 g payload (Pro) [published] (R16). Arctos $300–500, 2 kg, 600 mm, ±5 mm repeatability. Moveo $300–500 [reported] (conflicts: ~$600–800 / ~$1k elsewhere). Trade-off is repeatability: Arctos ±5 mm vs OpenManipulator-X <0.2 mm at $1,629.
Confirmed — budget motors, not plastic
Printing solves the cost problem
No. Actuators/compute dominate in every build: OpenManipulator-X servos $1,551.95 of $1,629.09 (95%); Poppy 60% actuators; ToddlerBot 90% computer+motors (R37). 3D printing removes the machine shop, not the motor bill.
Reframe budget
What this means for master84+: Keep the browser sim for logic, planning, and trajectory vetoes — that’s what it’s good at. Validate dynamics in MuJoCo before trusting a stacking threshold. Widen randomization (mass, friction, offset, lighting) instead of chasing prettier physics. And make the trajectory check a hard gate, not a warning — OT30 and master83 fail the same way when adjustment cases are waved through.
Dimension 1
Sim-to-real transfer: what breaks
Transfer works, never losslessly. The gap is dominated by contact / friction / tendon dynamics, actuator lag & gain, and out-of-distribution poses — not by rendering.
1.5 cm
Tobin 2017 (OT01) — sim-only RGB → real localization [published]. Foundational visual DR result.
11.5 vs 50
Dactyl (OT02) — median consecutive rotations, real vs sim [published]. ~100 years sim in 50 h on 6,144 cores.
60% / 20%
Rubik’s Cube (OT03) — typical / hardest scrambles solved [published] with ADR. ~13,000 years sim claimed [reported].
Locomotion & manipulation anchors
ANYmal speed
1.5 m/s
25% above prior 1.2 m/s record; 23–36% less torque than model-based control [published] (OT04). Fix was learning actuator dynamics from real data.
AME-2 parkour
2 m/s
Zero-shot on ANYmal-D + biped [reported] (OT21). Elevation-map error caps other systems at 0.6 m/s (OT06).
World-action zero-shot
35%
Real Franka average, no real demos (OT14) [published] vs 5% diffusion policy with 10 real demos. Far from deployable — but directionally real.
Learned actuator models from real data inside the sim (OT04).
DR-trained policies hold: 87–100% across DR levels vs 40% nominal (OT15, Sim2Cell / SO-101).
What doesn’t fix it: prettier graphics alone. R27: 11.5× lux drop barely moved success (−2.5 pts) because auto-exposure equalized brightness −3%.
Dimension 2
World models: from dream to policy
The lineage is clean: separate V-M-C → unified RSSM → decoder-free latent MPC → video-foundation world-action models.
Measured performance
Model
Result
World Models (OT10) Ha & Schmidhuber 2018
CarRacing 906±21 (first solve >900); VizDoom 1092±556 in dream [published]
DreamerV3 (OT09)
150+ tasks, one config; Minecraft diamonds from scratch [published]. Independent repro: humanoid 1049.99 vs ~193 random on T4 in ~3–4 h (OT49)
TT-MBRL (OT48)
Memory Maze 13×13: 46.2 → 185.0 with test-time adaptation [published]
Cosmos Policy (OT13)
LIBERO 98.5%, RoboCasa 67.1%, real ALOHA 93.6% [published]
GR00T 2 (OT41)
95% avg over 50 tasks after 5 demos/task [reported] — vendor, not peer-reviewed
3D representations — caution
Gaussian-splat MPC manipulates granular media with zero-shot generalization [reported] (OT34). SPLANNING gives rigorous collision-probability bounds in splats with a real 7-DoF demo [reported] (OT35).
Caveat (R34): Splats are rendering primitives, not collision geometry. Conversion to planner-safe geometry is required — renders don’t certify safe contact. Directly relevant to any splat-based world building.
DayDreamer (OT11): world models learn on real robots within hours [reported] — the clearest evidence imagination cuts real interaction time.
Dimension 3
Kit-bash costs: the motor bill is the bill
Source’s own numbers, with tags. Prices dated Nov 2026 or as stated in source and will move.
Build
Cost
Payload / note
EEZYbotARM (R39)
~$30 [reported]
Toy-class, SG90 servos; payload [unknown]
SO-101 pair (R16)
$259 kit / $229.88 parts [published]
500 g (Pro) [published]; 6 DoF, LeRobot-native; 11,788 datasets on HF (2026-11-01)
Aero Hand Open (R30)
$314 [published]
7 DoF, 16 joints, 389 g
Arctos 6-axis (R19)
$300–500 [published]
2 kg, 600 mm reach; ±5 mm repeatability
BCN3D Moveo (R20)
$300–500 [reported]
0.5 kg, 625 mm, ±0.1 mm — conflicts: ~$600–800 / ~$1k elsewhere
CRAFT hand (R32)
~$600 [reported]
15 active + 5 passive DoF
XLeRobot bimanual mobile (R24)
~$1,200 [published]
1.0 kg/arm; vs Mobile ALOHA ~$32k
RUKA hand (R31)
<$1,300 [published]
6 kg power grasp; 2.74 N pinch, 33 N slip
IRIS cinema arm (R22)
<$1,000 [published]
1.5 kg, ~1 mm repeatability, 940 mm reach
OpenManipulator-X (R23)
$1,629.09 [published]
500 g, <0.2 mm; servos = $1,551.95 (95% of cost)
LEAP Hand (R28)
$2,000 [published]
15 N/finger [reported]; 16 DoF, <4 h assembly
Berkeley Humanoid Lite (R37)
$4,312 US / $3,236 China [published]
0.8 m, walks
Poppy Humanoid (R37)
$8,000–9,000 [published]
60% of cost is actuators
Policy scaling on the cheap arm (R25)
On SO-101, ACT from scratch wins at scale: 1/40 @ 12 demos → 37/40 @ 100 demos[published]. π0.5 best when data is scarce (16/40 @ 12). SmolVLA never started in 47% of episodes. 1,340 scored episodes over 17 days, one task/arm/operator — one-task limit noted.
ACT vs Diffusion on SO-101 (R26)
Demonstrated positions: ACT 83.3% vs DP 81.7%. Interpolated positions: ACT 72.5% vs DP 52.5% (−10.8 vs −29.2 pts drop) [published]. ACT generalizes better between demos — relevant if you’re training from few demos.
Dimensions 4 & 5
Platforms & does sim-first actually save time?
Simulation platforms
Engine
Speed / note
MuJoCo
10–50× realtime / CPU core; cold start <1 s; best contact solver; only open-source differentiable engine (R05)
PyBullet
2–10× realtime; best for quick prototyping / RL flexibility (R12)
Isaac Sim / Gym
1–5× (GPU); thousands of envs/GPU; 4,096-env locomotion in <20 min (OT06)
Speed figures are rules of thumb [reported] (R05), not benchmark papers. Cost: MuJoCo/Gazebo/PyBullet/Webots/Genesis free (Apache/MIT); Isaac free-ish under NVIDIA EULA, needs NVIDIA GPU; Unity Industry $4,950/seat/yr (K09).
Time & cost saved — strongest evidence
RL training: 12–120 h → <20 min for quadruped locomotion [published] (OT06)
Virtual commissioning (OT27): ROI 3:1–5:1 [reported]; cycle-time prediction within 3–5%
Counter-evidence stays on the page: OT16 (sim misleads), OT03 (40% failure after ~13,000 years sim), OT14 (35% zero-shot), OT50 (DreamerV3 data-efficiency only weakly confirmed in traffic).
OT30 — Zero-net-wrench release. Basic stacking vs adjustment cases, 46 trials [published]. ATI Axia80 F/T sensor; gradient descent to zero external torque. Adjustment is where stacking dies — same pattern as master83.
—
OT29 — Contact-patch inference. Stability inferred from tactile + F/T during gentle contact on Bandu pieces [reported]. Quantitative accuracy in paper, not extracted [unknown]. Needs tactile + F/T sensing.
33 N
Nearest grasp anchors. RUKA slip threshold 33 N before fingertip slip [published] (R31) and Tobin 1.5 cm localization (OT01) are the closest quantitative anchors. No published grasp-offset threshold was found.
Implication: For stacking/placement, the literature offers methods (wrench-zeroing, contact-patch estimation) but not numbers to copy. Thresholds must be measured on your own rig — which is exactly what the pyramid runs have been doing.
Kaleidoscope — 10 cross-domain patterns
Why the pattern rhymes outside robotics
Biology: nervous system first is ancestral
Neurons & nerve nets predate brains by hundreds of millions of years; bilaterian ancestor ~700–650 Mya (K01) [published]. Conserved control architecture reused across wildly different bodies for ~600 My (K02, Arendt et al.) — the biological analogue of a portable control stack.
Film: kit-bashing as engineering
ILM industrialized kit-bashing; OED earliest “kitbash” citation 1972 (K03). Scarcity + scrap drove the “used future” — salvaged parts carry functional plausibility (K04). Greeble detail signals function (K05) — a loose parallel to why visual detail/randomization matters in sim-to-real.
Industry: sim-first beyond robots
PTC/Siemens synthesis: prototypes ~−50%, labor/prototype −26–30% [reported] (K06). One OEM: −10 months, −65% cost (K07). UE5 emerging as robotics sim platform for photorealism (K08) — vendor advocacy, no measured transfer numbers [unknown].
Browse the corpus
All 100 entries — grouped
Search or filter. Each entry carries its primary source in the research notes; relevance is direct / adjacent / distant as tagged by the researchers.
Could not verify — open questions (from the report, verbatim scope)
Browser-physics (Three.js / cannon.js) fidelity for robot control development — no primary source retrieved.
Published numeric thresholds for stacking stability, contact-aware placement, or grasp-offset measurement — methods exist (OT29–OT31), plug-in numbers do not.
Exact per-task numbers for OT05, OT18, OT20, OT24, OT39, OT42, OT43 — sources retrieved were summaries; figures exist in the papers but were not extracted.
All entries are index evidence (search-result text); none was verified live in the research browser. Prices are dated Nov 2026 or as stated and will move.