Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

1 Introduction

• ManuFacet-1K contributes 1,000 hours of force-synchronized demonstrations and deployment rollouts across three embodiments, two chassis families, and multiple manufacturing cells.The corpus adds synchronized wrist-wrench data to broad manipulation corpora.Source: §1 · p. 2
• Facet-0 aligns multi-view images and task instructions with kinematic state and causal wrist-wrench history, then jointly generates Cartesian action chunks and their expected future wrist wrench.Only the action is executed; wrench remains a predictive consequence.Source: §1 · pp. 2–3
• Deployment rollouts train a distributional Action–Wrench Critic and contact-selective credit to refine decisive interactions where similar geometric progress can conceal different contact outcomes.Controlled variants reached 16% success with semantic–contact alignment, 38% with value-guided refinement, and 82% with bounded local adaptation. Against advantage-weighted behavior cloning, intervention fell from 47% to 24% and failure recovery rose from 44% to 81%.Source: §1 · pp. 2–3
• Bounded local adaptation reuses the frozen representation to produce executable Cartesian targets under changed part dynamics while retaining predictive action–contact coupling.With ten demonstrations and three hours of training on 6.6% of parameters, adaptation reached 45% success on an unseen module versus 5% for the strongest baseline.Source: §1 · p. 2
• 82% mean success across five precision computer-assembly tasks versus 15% for the strongest matched RECAP-style baseline, with 0.5 mm placement accuracy and 50 ms command latency.The full model reaches these results across five sub-millimeter tasks.Source: §1 · pp. 2–3

2.1 Vision-Language-Action Foundation Models

• Modern VLA models span autoregressive, diffusion, and flow-matching action chunks, while Facet-0 fuses causal wrench history with language and kinematics to generate actions with anticipated wrench.This joint proposal makes interaction quality available to value-guided post-training, rather than treating contact only as an observation.Source: §2.1 · p. 3

2.2 Force- and Tactile-Aware Policies

• ForceVLA [49], TA-VLA, Tactile-VLA [22], FoAR [19], FD-VLA, and ForceFlow [51] [53] use contact signals, whereas Facet-0 values anticipated wrench jointly with proposed motion.Prior approaches introduce force as an input, auxiliary target, or local control cue; Facet-0 uses the joint proposal to shape which interactions policy improvement favors.Source: §2.2 · p. 3

2.3 Contact Control and Compliant Assembly

• Hybrid force-position, impedance, variable-impedance, and energy-tank control regulate contact [5] [1], while Facet-0 learns to anticipate and value contact consequences.Classical transfer across workpieces can require task-frame identification and parameter retuning; compliant control and explicit workspace and per-step motion limits remain at Facet-0's execution layer.Source: §2.3 · p. 3

2.4 Learning from Deployment Experience

• Real-robot RL spans correction-guided learning [31] [33], value-driven intervention reduction [12] [11], deployment pipelines [28], VLA post-training [30], and frozen-policy adaptation [47] [3] [24] [43] [9].Facet-0's Action–Wrench Critic scores predicted action-wrench pairs, and policy improvement in action–wrench space changes behavior during contact; intervention rate is reported as a metric.Source: §2.4 · pp. 3–4

2.5 Manipulation Datasets

• Broad datasets prioritize behavioral coverage [25] [2] [48], while REASSEMBLE and FMB offer smaller-scale assembly resources [32] [44]; ManuFacet-1K complements both with force-synchronized, phase-annotated data.It spans three embodiments, two chassis families, and multiple centers under a shared hardware and teleoperation specification; its skill-phase taxonomy supports reward decomposition and refinement conditioning.Source: §2.5 · p. 4

3.1 Corpus Composition

• The ~1,000-hour corpus covers CPU (37.3%), RAM (21.9%), Disk (23.4%), and GPU (17.3%) installation, captured with wrist force/torque sensing and three RGB cameras.The curated data combines open-loop teleoperated demonstrations with closed-loop policy rollouts involving human intervention. Each frame stores a 13-D state; action resolution is 0.5 mm, and VLM-generated sub-task/instruction pairs accompany segments.Source: §3.1 · p. 5

3.2 Skill-Phase Taxonomy and Statistics

• Seven manufacturing phases are defined, with align, insert, press, seat, and fasten designated critical because precision determines outcomes.Contact-phase frames are a minority but carry most outcome information; peak forces are bimodal, with a seating mode roughly an order of magnitude above low-force alignment, depending on the part. Boundaries are marked during collection and verified offline, with disagreements re-annotated.Source: §3.2 · p. 5

3.3 Collection and Curation

• Curation synchronizes multimodal streams, extracts valid segments, samples salient action chunks, adds VLM-assisted sub-task and instruction supervision, and corrects anomalies while retaining corrected trajectories.Full rates, thresholds, annotator protocol, and reflow rules are deferred to Appendix B.Source: §3.3 · p. 5

4.1 Multimodal Grounding

• The multimodal policy fuses vision-language features, kinematic state, and a 10-step wrist-wrench history to represent contact as temporal events.Its inputs include three synchronized RGB views, language, end-effector pose, and gripper opening; the wrench history can encode contact onset, persistent off-axis moment, and rising wrench under stalled motion.Source: §4.1 · p. 6

4 Semantic–Contact Alignment

• The policy jointly proposes Cartesian actions and their anticipated wrist-wrench consequences, but only the actions are executable commands.The 50-step action–wrench chunk pairs each action with a predicted subsequent wrench; coarse proposals are refined without changing this joint parameterization. A structured attention mask blocks cross-attention between action–wrench and VQA targets, while robot state enters only the action expert.Source: §4 · p. 6, §4.1 · p. 7

4.2 Predictive Interaction Learning

• Conditional flow matching trains joint action–wrench prediction, while a separate VQA objective supervises current-subtask and next-instruction answers.At inference, reverse integration generates the joint proposal, so reducing wrench prediction error requires encoding how proposed motion changes contact; separate optimizers alternate the two objectives.Source: §4.2 · p. 7

5.1 Interaction Valuation

• A distributional Action–Wrench Critic ranks joint action–wrench proposals using phase-aware sub-goal progress and contact rewards, distinguishing clean insertions from jams at equal depth.A vision-language judge segments rollouts into sub-goals and labels contact phase; distributional prediction preserves distinct success and failure outcomes rather than averaging them.Source: §5.1 · pp. 7–8, Figure 4 · p. 8

5.2 Contact-Selective Credit

• Contact-selective credit labels top-ranked frames separately within contact and free-space regimes, then routes positive tags to guided refinement while training the unconditional branch on all frames.Short-horizon credit is weighted by contact intensity in the refinement loss, concentrating policy improvement on decisive interactions.Source: §5.2 · pp. 8–9

5.3 Local Policy Adaptation

• For part-specific transfer, a lightweight actor reuses frozen representations and a policy reference but is trained and valued only over bounded executable Cartesian actions.Its auxiliary wrench head predicts the next measured wrench but is neither an action dimension nor a force command; only the bounded action is executed, with force control using wrist feedback.Source: §5.3 · p. 9

6.1 Experimental Setup

• The assembly suite covers five host-computer installation tasks, 23 sub-goals, and 20 trials per cell under a pre-declared protocol with no re-runs.The suite includes nine free-space sub-goals and fourteen contact-critical sub-goals at 0.10–0.30 mm clearance, with randomized chassis pose and per-phase safety limits.Source: §6.1 · p. 10

6.2 Main Comparison and Component Gains (Q1)

• Facet-0 achieves 82% mean task success versus 15% for the strongest baseline, 𝜋0.5+RECAP-style, a 5.5× improvement.The controlled variants progress from 16% for Align to 38% with value-guided RL and 82% for Full. Full reaches 85–95% on four insertion tasks but 50% on LEVER; it issues commands every 50 ms versus 150 ms for 𝜋0.5 and completes tasks 40% faster than an expert teleoperator.Source: Table 2 · p. 11, §6.2 · pp. 11–12

6.3 Post-Training from Deployment Experience (Q2)

• On the Disk deployment corpus, contact-selective post-training raises success from 20% to 65% and recovery from 44% to 81%, while intervention falls from 47% to 24%.Unlike scalar-advantage AWR, contact-selective credit ranks samples within interaction regimes, retaining high-credit contact-recovery frames that are scarce in demonstrations.Source: §6.3 · p. 12, Figure 7 · p. 12

6.4 Adaptation: Paradigms and Few-Shot Transfer (Q3)

• Across adaptation episodes, DSRL stays at the floor while both Facet-0 variants reach ceiling success; the TD3+BC actor–critic ends at full success on both tasks.The imitation variant starts high and remains comparatively flat; the actor–critic variant starts lower, dips on RAM insert, then recovers. Autonomy rises for both. Against position-only residuals, the bounded actor also predicts wrench consequences, while the wrench head remains non-commanded and the refinement expert stays frozen.Source: §6.4 · pp. 13–14, Figure 8 · p. 13
• On an unseen memory module, Facet-0 reaches 9× the strongest baseline while updating a fifteenth of the policy parameters.All methods receive ten demonstrations and three hours of training; Facet-0 trains only the adaptation actor for the changed part dynamics.Source: §6.4 · pp. 13–14, Table 3 · p. 14

6.5 Ablation Studies (Q4)

• On RAM insertion, jointly removing wrench prediction, valuation, and adaptation cuts success by 50 points and raises peak off-axis force from 0.2× to 1.0×.Individual removals differ by only one to two successes in 20 trials, so their ordering is unresolved; removing critic or adaptation wrench raises peak force to 0.5× or 0.4×, while removing predictive wrench preserves 0.2×.Source: Table 4 · p. 14, §6.5 · p. 14

7 Discussion and Limitations

• Controlled evaluation covers five wrist-sensed, parallel-gripper electronics-assembly tasks on a shared chassis fixture, despite a corpus spanning three embodiments and two chassis families.Within this setting, results support semantic–contact representation learning, value-guided refinement, and bounded adaptation; the contact-role ablation distinguishes prediction, valuation, and adaptation.Source: §7 · p. 14
• Future work will broaden controlled evaluation to more end effectors, parts, and contact geometries, and explore force-conditioned policy distillation for cells without wrist-mounted six-axis sensing [53].Source: §7 · p. 14

8 Conclusion

• 82% mean success on the five-task electronics-assembly suite (15% for the strongest baseline), with 0.5 mm placement accuracy and 50 ms command latency.The full model combines force-synchronized data, joint action–wrench prediction, an Action–Wrench Critic, and bounded local adaptation.Source: §8 · p. 14
• Controlled variants reach 16% success with semantic–contact alignment, 38% after value-guided refinement, and 82% with local adaptation.Source: §8 · p. 14
• Against the matched AWR baseline, value-guided post-training reduces intervention from 47% to 24% and increases recovery from 44% to 81%.Source: §8 · p. 14
• The Action–Wrench Critic supports global refinement, while a bounded actor reuses the frozen representation and policy reference for unseen-module transfer.On wrist-sensed electronics assembly, the actor adapts with ten demonstrations while updating 6.6% of parameters; extension to further embodiments and contact regimes is ongoing, and the model, dataset, and training recipe are released.Source: §8 · p. 15

A Assembly Suite Details

• The assembly suite uses a shared chassis fixture with distinct contact geometries, and LEVER is evaluated as isolated contact-only CPU socket lever closure.It is not evaluated as another full pick-to-place chain; Fig. 5 records execution chains, while Table 5 gives clearances and configured force limits.Source: A Assembly Suite Details · p. 19
• The host-computer suite comprises 23 sub-tasks, 14 contact-critical; CPU force limits are 8 N for pin-array seating and 35 N for socket closure.Clearance denotes the tightest fit among a task’s critical sub-goals.Source: Table 5 · p. 19

A.1 Evaluation Protocol

• Five tasks span 23 sub-tasks and seven shared verbs; nine free-space sub-tasks use coarse chunks, while 14 contact-critical sub-tasks activate refinement at 0.10–0.30 mm clearances.Each contact phase also has a configured safety limit Fmax.Source: §A.1 · p. 19
• Evaluation randomizes chassis pose, uses feeder-supplied parts, and fixes each main-table cell at 20 no-rerun trials under a predeclared protocol.Metrics are success rate (completion without force violation), peak off-axis force, violation rate, intervention rate, recovery rate, cycle time, and inference latency.Source: §A.1 · p. 19
• Baselines include π0.5, π0.5+F, aligned π0.5+RECAP-style, GR00T N1.7, and aligned TA-VLA, trained and evaluated under matched budgets and one protocol.π0.5+F adds a wrench input token; TA-VLA aggregates wrist force/torque history into one token.Source: §A.1 · p. 19
• Facet-0 Align denotes semantic–contact alignment, +RL adds value-guided refinement, and Full additionally includes local policy adaptation.The protocol identifies π0.5+RECAP-style as an aligned reimplementation, not the released π★0.6 model.Source: §A.1 · p. 19

B Dataset Curation Protocol

• Each acquisition cell uses a torque-controlled arm, parallel gripper, wrist six-axis force/torque sensor, three RGB cameras, joint proprioception, and force-feedback kinesthetic teleoperation.Hardware timestamps align a shared 15 Hz training timeline, while high-rate state and wrench streams are retained for the 200 Hz controller.Source: B Dataset Curation Protocol · pp. 19–20
• Curation rejects episodes with excessive cross-channel skew and removes faulty or corrupted segments while retaining unaffected intervals; action and wrench variation guides down-sampling of pauses and waits.Removed faults include teleoperation errors, truncated chunks, sensor saturation, calibration drift, and sustained part occlusion.Source: B Dataset Curation Protocol · p. 20
• Segments receive VLM-generated task and next-instruction pairs, seven-phase boundary annotations verified by a VLM, and corrected-trajectory pairings that preserve anomaly and recovery signals.The phase labels are reused for pretraining and reward decomposition; center identities are pseudonymized, identifying content removed, and the release includes a data card and curation code.Source: B Dataset Curation Protocol · p. 20

Why the wrench history is a window

• A 10-frame wrench window captures contact onset above sensor noise, while jams and skews require short trajectories rather than single samples.These contact patterns may no longer be visible to vision after a part occludes the slot.Source: C Architecture and Execution Details · p. 20

Why the wrench is predicted rather than only read

• Predicting wrench makes the model represent the contact consequence of proposed motion before execution, unlike conditioning on wrench produced by past actions.The wrench loss rewards representing that consequence, making the predicted wrench available before execution.Source: C Architecture and Execution Details · p. 20

Execution hierarchy

• A joint action–wrench proposal decoded at 5–10 Hz is refined to 20 Hz at roughly 0.5 mm by a contact-rate expert.The proposal is decoded at roughly 5 mm resolution; coarse and fine actions differ in precision, not kind. The refinement expert is frozen during local policy adaptation.Source: C Architecture and Execution Details · p. 20

Compliant controller

• The 200 Hz compliant controller closes its force loop on measured wrist feedback, using predicted wrench only for representation and value learning.It regulates force along sub-goal-designated task-frame directions and tracks position along the remainder [5]; predicted wrench is neither a controller input nor a force setpoint.Source: C Architecture and Execution Details · p. 21

Safety operator

• Stask clips targets to the workspace, limits per-step motion, handles rotation, optionally smooths and gates commands, and passes gripper commands unchanged.Its displacement bound limits tool travel between observations but is not a passivity guarantee; no certified hard force bound is claimed.Source: C Architecture and Execution Details · p. 21

Why adapt behind an interface rather than fine-tune the backbone

• Local adaptation updates only a bounded task actor, preserving the frozen semantic backbone and constraining every output to the admissible action box before Stask.The actor adapts to an unseen part in hours on one robot; the absolute bounded parameterization avoids relying on small residual corrections to repair base-policy failures.Source: C Architecture and Execution Details · p. 21

Predictive coupling across optimization

• During local adaptation, a supervised one-step wrench head preserves predictive action–wrench coupling, while twin critics and the robot interface operate only on executable Cartesian actions.The wrench prediction remains auxiliary and non-commanded.Source: C Architecture and Execution Details · p. 21

D Method Details

• The three method modules reuse the semantic–contact representation: alignment decodes joint action–wrench proposals, refinement scores them with the Action–Wrench Critic, and local adaptation compresses the representation while adapting executable actions.The predicted wrench remains coupled to action consequences rather than becoming a commanded action.Source: D Method Details · p. 21

D.1 Semantic–Contact Alignment

• Alignment masks invalid action and wrench timesteps separately, alternates action–wrench and VQA updates four-to-one, and over-samples precision and recovery episodes relative to their duration.Separate mask normalization prevents batches with few valid contact frames from silently reducing the force-term contribution; the run uses 32K action-branch and 8K VQA updates.Source: §D.1 · p. 22
• Each action row in the joint proposal is paired with the measured wrench one step later, assigning contact consequence to the motion that produced it while masking invalid future samples.The offset does not change the action horizon or number of executable commands.Source: §D.1 · p. 22

D.1.1 Held-out Alignment Diagnostics

• The joint VQA-plus-wrench model reduced answer-token cross-entropy by roughly three orders of magnitude while reaching nearly the same action–wrench loss endpoint as wrench-only training.The comparison used a matched budget of 32K action-branch updates; validation losses are not task-success measures.Source: Figure 11 · p. 23, §D.1.1 · p. 23

D.2 Value-Guided Policy Refinement

• The distributional Action–Wrench Critic uses auxiliary heads for near-future wrench, contact intensity, within-contact progress, and success ranking to distinguish equal-progress states with different contact outcomes.During refinement, action-to-next-wrench supervision is retained with causal filtering, per-axis normalization, and contact-validity masking; executable action columns remain unchanged.Source: §D.2 · p. 24
• Refinement selects positive interaction labels by contact regime and matched overall positive rate, then conditions the policy with a literal instruction tag; contact intensity uses causally smoothed wrench differences.The unconditional branch continues to see the full data distribution, and guidance scale zero recovers the unconditional policy.Source: §D.2 · p. 24

D.3 Local Policy Adaptation

• Local adaptation freezes the semantic–contact encoder, bottleneck, reference policy, and other upstream components while a lightweight TD3-style actor–critic learns bounded Cartesian targets for the new part.The FACET token, proprioception, measured wrench, and frozen reference action inform the actor; successful human interventions provide action anchors, and later confirmation filters out failed intervention segments.Source: §D.3 · pp. 25–26
• Only the action head controls execution, while the auxiliary wrench head predicts measured next-step wrench; critics learn from the safety-filtered action actually applied to the robot.The wrench head regularizes the shared trunk without adding a force-command interface, and replay uses the applied command because it caused the observed reward and next state.Source: §D.3 · p. 26

• Table 7 reports 17 of the suite’s 23 sub-goals, including 13 contact-critical rows, as a diagnostic subset rather than a complete decomposition.The displayed subset is also visualized for representative methods in Fig. 6.Source: E Reported Per-Sub-Goal Results · p. 27, Table 7 · p. 27
• Each full-system chain rate matches its weakest displayed contact-critical sub-goal: alignment limits RAM, CPU, Disk, and GPU, while the first press limits LEVER.Table 7 reports the sub-goal decomposition for 20 trials per cell.Source: Table 7 · p. 27, E Reported Per-Sub-Goal Results · p. 27
• The 13-row contact mean weights shown contact events equally, while the five-task mean weights assemblies equally to measure end-to-end reliability.The contact mean is rounded to the nearest point; task rows require completion of the full chain, whereas sub-goal rows isolate individual decision points.Source: E Reported Per-Sub-Goal Results · p. 27

F Supporting Evaluation Data

• Facet-0 completes the task 40% faster than expert teleoperation on the same fixtures.The section retains the exact reference values behind the main-text visualizations.Source: F Supporting Evaluation Data · p. 28
• Table 8 covers deployment efficiency and precision, measuring placement accuracy at the align-to-insert transition.The supplied passage gives the table scope but no metric values.Source: Table 8 · p. 28
• Table 9 reports disk post-training results underlying Fig. 7, with matched deployment data, VLM reward, and budget.The supplied passage identifies the comparison scope but gives no result values.Source: Table 9 · p. 28
• Table 10 gives adaptation reference values for a two-hour window and 30-minute checkpoint; violations cover all training episodes, and a dash means 90% success was not reached.The listed measures include time to 90%, success at 30 minutes, and violations.Source: Table 10 · p. 28, F Supporting Evaluation Data · p. 28

F.1 Behavioral Analysis

• On RAM and CPU, the adapted actor uses off-axis wrench feedback to retreat, correct laterally, and re-enter, unlike π0.5 continuing its downward chunk.The correction remains within the same sub-goal and runs at the contact-control rate, without requiring backbone task replanning.Source: §F.1 · p. 28
• During press-and-lock, wrist force distinguishes a jam from successful seating when gripper occlusion hides misseating: force rises while displacement stalls.The system can flag the jam, re-issue the sub-goal, and complete insertion.Source: §F.1 · p. 28
• Per-sub-task Eq. (4) credits a recovered trajectory’s corrective segment, while regime-conditioned selection retains rare contact-recovery frames in the positive set.Together, these mechanisms improve tail behavior governing whether a failed interaction resumes or halts the cell.Source: §F.1 · p. 28
• Table 7 measures complete chains and contact decisions, Table 9 isolates deployment post-training changes, and Table 10 records local-adaptation speed and safety at separate operating scales.Keeping their denominators separate connects task completion to recovery and adaptation without combining distinct outcomes into one aggregate score.Source: §F.1 · p. 28