A CMU preprint updated to v2 on arXiv (October 6, 2026) presents SAPS: Shared Autonomy for Policy Steering by Blending Teleoperation with a Pretrained VLA (arXiv:2606.15568v2, cs.RO, funded by NSF Grant No. 2341352). The headline result in one sentence: full robot autonomy succeeds 8.3% of the time, pure human teleoperation 34.5% on the real robot, and human-robot shared control up to 77.4%.
How SAPS Arbitrates
VLA models are brittle: shifts in object position, appearance, or instructions can derail the policy. Prior fixes add dynamics models, modify diffusion samplers, or attach explicit reasoning modules. SAPS adds nothing — it arbitrates only at the action level, blending human joystick commands with the π0.5 policy's actions (15 Hz on the real robot, one decision per step).
Three arbitration modes, differing only in how α is computed:
- Takeover: any human input (L2 norm > ε = 10⁻³) sets α to 0 (pure human control); when input stops, α returns to 1.
- Blending: human input sets α to a fixed 0.5 (50/50); returns to 1.0 otherwise.
- Cosine: compute cosine similarity c between human and policy actions, then γ = 1/(1+e^(−6c)), α = γ. The more aligned the directions, the more weight the policy gets.
- Blending equals teleoperation in simulation (44.0%, p = 0.92) — a fixed 50/50 mix added zero gain on sim tasks.
- On the real robot, all three shared autonomy methods significantly outperform teleoperation (Wilcoxon signed-rank, p ≤ 0.017); in simulation only Cosine is significant (p = 0.015).
- Versus full autonomy, all comparisons pass Holm-corrected Fisher exact tests (all p < 0.005).
- Real-robot intervention rates cluster tightly at 43–45% across methods.
Gripper control is latched separately: users can switch gripper state at will, while the policy must toggle twice in simulation or wait a 5-second timeout on hardware — a design giving users more grasp control and fewer accidental drops.
Results
14 users (ages 18–37, 9 male, 5 female) ran LIBERO-PRO in simulation; 13 ran the real-robot portion (one withdrew due to hardware setup difficulty). Each task/method was repeated 3 times, 12 episodes per task.
| Metric | Full Autonomy | Teleop | Cosine | Blending | Takeover | |---|---|---|---|---|---| | Sim success rate (%) | 8.3 | 44.0 | 75.0 | 44.0 | 65.5 | | Real success rate (%) | 8.3 | 34.5 | 67.2 | 64.3 | 77.4 | | Sim completion time (s) | 13.3 | 23.2 ± 3.1 | 20.3 ± 5.2 | 17.8 ± 4.2 | 21.4 ± 5.3 | | Real completion time (s) | 44.8 | 42.4 ± 4.5 | 36.1 ± 7.9 | 34.4 ± 6.9 | 37.8 ± 7.3 | | Sim human intervention (%) | 0.0 | 100.0 | 42.9 ± 14.2 | 46.5 ± 20.6 | 32.2 ± 10.3 | | Real human intervention (%) | 0.0 | 100.0 | 43.0 ± 12.4 | 45.4 ± 13.8 | 44.3 ± 10.2 |
Notable points:
Workload (NASA-TLX)
| Method | Sim TLX (n=14) | Real TLX (n=13) | |---|---|---| | Teleop | 49.0 (18.6) | 53.8 (18.7) | | Takeover | 32.2 (13.6) | 36.2 (18.6) | | Blending | 53.1 (18.0) | 38.5 (18.4) | | Cosine | 31.1 (18.2) | 41.2 (11.7) |
Simulation's best is Cosine (31.1, ~37% lower than teleoperation); the real robot's best is Takeover (36.2, ~33% lower). Blending in simulation scores *worse* than pure teleoperation (53.1 vs 49.0).
Case Study: A Participant with Tetraplegia
One participant, a 40-year-old male with complete ASIA C5 tetraplegia (injury 20 years prior), used an Xbox Adaptive Controller driven by residual shoulder movement. Results were mixed: in LIBERO his success rate improved 8.3% and completion time dropped 16.9%; on hardware, time dropped 15.5% but success matched teleoperation. Despite lower reported workload for shared methods in simulation, he ranked teleoperation first in a forced-choice ranking — consistent with prior assistive shared autonomy findings that users may prefer retaining control even at the cost of objective performance.
Stated Limitations
The authors are direct: SAPS requires a human in the loop at test time and depends on operator timing, skill, and the teleoperation interface. Because it blends with the policy rather than estimating task progress, intent, uncertainty, contact, or safety, it fails when the policy is "confidently wrong." Cosine and Blending can drive objects into obstacles when human and policy intent diverge; the authors suggest more dynamic arbitration for such cases. Evaluation covered a limited task set, operators, embodiments, and policy backbones.
Why It Matters
The dominant embodied-AI narrative is making models run autonomously for longer. SAPS offers an alternative: when the model stalls, a human steps in — and those seconds count as takeover, not failure. Deployment cost is low: π0.5 is an open-source pretrained VLA, and SAPS arbitrates its output actions without architectural changes, retraining, or auxiliary models.
Caveats: the paper is a non-peer-reviewed preprint (v1 submitted 2026-06-14, v2 revised 2026-10-06). Author order was decided by coin flip: Crystal Zhou and Jehan Yang are co-first authors; Douglas J. Weber and Zackory Erickson are co-senior authors. Project page: shared-autonomy-policy.github.io. The case study is a single individual, not group statistics.