Dual RGB + language + proprioception
Frozen RDP trunk
Produces four slow visual action proposals. Its action is used only in the visual phase.
slow action proposal
Supported
The frozen hierarchy/controller treatment improves goal success over the registered RDP and DP comparators on this protocol.
Not supported
The evidence does not show tactile necessity, tactile-ratio robustness, safe success, multi-seed robustness, or broad task generality.
01 / paired confirmation
The hierarchy ordering is statistically supported
Object-soft tasks 0-9, resets 5-9, simulator seed 2000, checkpoint seed 0. All methods used the same 50 reset states, horizon, scorer, initialization protocol, and honest runtime observations.
Oursdecentralized RDP hierarchy
Wilson 95% CI [44.2, 70.6]
Modified RDPregistered comparator
Wilson 95% CI [14.3, 37.4]
Frozen DPregistered comparator
Wilson 95% CI [0.0, 7.1]
Integrity passed.
All 50 reset-state hashes matched across the six confirmation arms.
One registered infrastructure retry was replaced with an exact matching cell.
02 / causal follow-up
Touch is active, but not load-bearing here
The preregistered follow-up reused the 50 confirmation cells. It intervened on sensor availability and tactile observer training while leaving the hierarchy and controller fixed.
Tactile-necessity gate
FAIL
Runtime touch-off exactly tied normal ours. The familywise one-sided 95% lower bound was -14 points; every Holm-adjusted one-sided paired p-value was 1.000.
| Arm | Goal success | Wilson 95% CI | Learned contact observed | Controller complete | Safety unknown |
|---|---|---|---|---|---|
| 100%-trained, normal touch | 29/50 · 58% | [44.2, 70.6] | 27/50 | 35/50 | 8/50 |
| 100%-trained, runtime touch off | 29/50 · 58% | [44.2, 70.6] | 0/50 | 37/50 | 9/50 |
| 12.5%-trained, normal touch | 30/50 · 60% | [46.2, 72.4] | 26/50 | 38/50 | 8/50 |
| 0%-data, untrained observer | 30/50 · 60% | [46.2, 72.4] | 0/50 | 36/50 | 9/50 |
What the intervention isolates
Runtime-off records sensor_gate=[false,false] and disables tactile
messages at every inference. It is the direct no-sensing intervention.
Why performance survives
RGB/proprioceptive preemption, task pose primitives, and a fixed 24-step maximum close-hold fallback can advance transport without tactile authorization.
Correct benchmark claim
Keep this as evidence for hierarchical control on object-soft tasks. Do not present its method gap as evidence that tactile sensing caused the gain.
03 / method
Role-specific observers, one causal authority
Observation representations are partly decentralized. Each tactile side exports a compact contact vote, while one central phase controller chooses the final 7D action.
Left marker-flow history
Left encoder + FSQ
Private tactile representation; exports contact and persistence probabilities.
left compact voteRight marker-flow history
Right encoder + FSQ
Independent weights and input history; never receives the left marker stream.
right compact voteRobot pose + task identity
State gate + pose primitive
Preempts vision near the grasp and supplies close and transport poses.
phase preemptionArchitecture boundary: this is role-specific observation with centralized arbitration. It is partial decentralization, not a multi-agent policy system.
04 / clean rollout evidence
Native simulator camera and tactile streams
Direct task-4, reset-0 streams from the performance renderer with TAA. Each file is displayed at or below its native resolution with no dashboard compositing, denoising, temporal filtering, or web re-encoding.
Render quality passed
Median raw-frame speckle 0.087%
Reject threshold 3.0%
512 px RGB · 320x240 tactile · 10 fps
Ours · successful rolloutTask 4 / golden striped pastry
Direct simulator capture
Fixed camera · native 512 px
Private observer 02LLeft marker flow
Native 320x240
Private observer 02RRight marker flow
Native 320x240
All three players are linked on the same 10 Hz simulator timeline.
Evidence boundary: this is a clean rerender of the earlier development task-4 method cell, not one of the 50 confirmation cells. It visualizes the controller; the confirmation statistics remain the tables above.
Annotated composite views
These legacy dashboards enlarge native frames to make phase labels and three-arm comparisons readable. Use the direct streams above for visual inspection.
Method walkthroughFixed camera + wrist + bilateral tactile
Scaled annotation view
Paired comparisonOurs vs modified RDP vs frozen DP
5 tasks · reset 0
05 / cross-benchmark context
TacEx Factory Wipe rollout
Restored historical camera evidence from a matched Wipe state. These clips use an earlier method revision and are not comparable with the SoftVT confirmation rates.
Policy perception
Onboard proprioception + bilateral force history
RGB is simulator visualization only
Seed 2000 · 12.0 s · 24 fps
Historical comparatorDP
Native 224x112
Historical comparatorRDP
Native 224x112
Earlier method revisionSlow-fast late fusion
Native 224x112
Evidence boundary: this is task visualization from an earlier matched development study. It is not footage from the current SoftVT method, not a TacEx confirmation cell, and not evidence for a current DP/RDP/ours ordering. TacEx records force-history vectors rather than tactile RGB, so there is no corresponding tactile image video in this artifact.
06 / development artifact
Original five-cell benchmark overview
This frozen image reports the original development result: ours 4/5, RDP 2/5, DP 0/5. Its task-4 storyboard now uses the clean performance-renderer replay; the original noisy image remains available below.
07 / repository map
Benchmark and simulation atlas
The former standalone atlas is preserved here as a dated repository snapshot. It maps the live, external, and retired benchmark lanes that existed on 2026-08-12.
Archived source
RoboProgram Benchmark & Simulation Atlas
Browse adapter locations, basic evaluation commands, baseline context, and the original paper-lane recommendation without mixing those historical repository notes into the SoftVT or TacEx result claims above.
Snapshot boundary: this atlas is retained for provenance and may not match current main. Its cross-benchmark rows use different protocols and remain context only; they are not pooled with or directly ranked against the SoftVT and TacEx studies on this page.
08 / provenance
Scope and reproducibility
The downloadable manifest records revisions, hashes, protocols, and media sources.
| Perception regime | Honest runtime dual RGB, language, proprioception, and registered tactile treatment |
|---|---|
| Confirmation protocol | Tasks 0-9, resets 5-9, simulator seed 2000, checkpoint seed 0, 50 paired cells per arm |
| Ordering revisions | 7974b1c evaluation / 3799137 analysis |
| Necessity revisions | 1ed8789 evaluation / 0133165 analysis |
| SoftVTBench revision | 5805611 |
| TacEx rollout scope | Historical matched development visualization; onboard proprioception + force history; camera excluded from policy input |
| Safety | Not certified; 35/300 ordering cells and 34/200 ablation cells have unknown deformation safety |
| Generalization boundary | One checkpoint seed and one object-soft suite; no multi-seed or cross-suite claim |