Source-linked AI summary
RoboPace: Contact-Aware Time-Optimal Retiming for Action-Chunk Policies
Mimo Shirasaka, Takehiko Ohkawa, Takuya Okubo, Nicola Scianca, Tatsuya Matsushima, Kei Ota
TL;DR
Demonstration timing can mismatch a robot’s capabilities: it may be too fast near contact and too slow in free space. RoboPace retimes policy-predicted paths online using predicted contact and robot constraints, without retraining; across three contact-rich tasks, it achieved higher overall success than slow uniform execution while completing four of five commands in about half the time.
Problem
Demonstrations encode the demonstrator’s timing, which may be too fast for a robot near contact yet slower than its available free-space speed.
Source: §1 · p. 2Method
Results
64% overall success versus 44% for slow uniform execution, with four of five commands completed in about half its time across three contact-rich tasks.
Source: §1 · p. 2, §6 · p. 11Takeaways & Limitations
RoboPace adapts execution timing online without policy retraining while preserving the policy’s predicted path.
Source: §1 · p. 2, §6 · p. 11The evaluation covers one robot and one policy family; generalization across robots, policies, tasks, and prediction horizons remains future work.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Robot manipulation data collection has been shifting from teleoperation toward robot-free demonstrations, through interfaces such as the Universal Manipulation Interface (UMI) or directly from human hands. Vision-Language-Action (VLA) policies trained on such data inherit the demonstrator's timing. Yet human timing does not directly transfer to robots: compliant hands tolerate fast contact, whereas robots may overshoot due to actuator and tracking limitations; conversely, robots can move faster in free space. This motivates a unified approach that reconciles execution speed with contact safety. We present RoboPace, an online retiming layer that preserves the policy's geometric path while adapting its timing, respecting the target robot's kinematic and dynamic constraints. It adapts execution speed based on predicted contact, jointly accounting for contact-dependent speed limits and the robot's motion constraints. The method requires no policy retraining and operates in real time. Across three contact-rich tasks on a dual-arm robot, faster uniform execution and physical-limit-only retiming largely fail. RoboPace instead achieves higher overall success than slow uniform execution while completing four of five commands in approximately half the time, retaining the reliability of slow execution without its time cost.
1. Introduction
Demonstration timing can be too fast near contact yet unnecessarily slow in free space, a trade-off that a global speed factor cannot resolve. RoboPace retimes action chunks according to predicted contact and robot constraints; across five commands, it outperformed slow uniform execution in overall success while completing four commands in about half the time.
1. Introduction
- Demonstration timing can be too fast near contact for a robot’s compliance, actuator limits, and tracking dynamics, yet slower than the robot can move in free space.Slowing the whole path gives up free-space speed, while speeding it up risks failure near contact.Source: §1 · p. 2
- Prior methods do not jointly impose predicted-contact speed limits and the target robot’s kinematic and dynamic constraints during retiming.Training-time approaches may require changing demonstrations or retraining, while cited inference-time approaches address robot limits or criticality-based rate switching separately.Source: §1 · p. 2
- RoboPace applies a position-dependent velocity ceiling near predicted contact and combines it with robot constraints in online time-optimal retiming, without retraining the policy.Geometric proximity and wrist-camera predictions can each tighten the ceiling, with smooth transitions back to full speed away from contact.Source: §1 · p. 2
- 64% overall success versus 44% for slow uniform execution, with four of five commands completed in about half its time.The evaluation covered three contact-rich tasks on a dual-arm robot.Source: §1 · p. 2
2. Related Work
The section reviews approaches to demonstration timing and inference-time speed adaptation, then positions RoboPace as an integration of contact-aware speed modulation with robot-constrained retiming. RoboPace leaves the policy and its predicted geometry unchanged and requires no policy retraining.
2. Related Work
- Imitation-learned policies inherit demonstration timing, including VLA and earlier policies, while demonstrations increasingly span robot and robot-free collection settings.Action chunks are learned from end-effector trajectories and executed on a target robot.Source: §2 · p. 2
2.1. Training-time speed adaptation
- Training-time methods alter demonstration timing, policy timing, objectives, or action representations to adapt execution speed.Examples include entropy- or semantics-based down-sampling, tempo-conditioned policies, stage-adaptive speeds, and temporally scaled continuous curves.Source: §2.1 · p. 3
2.2. Inference-time speed adaptation
- TOPP computes fast timing for a fixed path under robot constraints, and established methods also impose Cartesian velocity bounds or proximity-based speed reduction.TOPP solvers include convex optimization, numerical integration, and reachability analysis.Source: §2.2 · p. 3
- RACE retimes action chunks under robot constraints without a predicted-contact path constraint, while SAIL switches among preset rates rather than solving constrained path parameterization.Other cited inference-time methods address chunk consistency, execution horizon, or procedural-phase routing.Source: §2.2 · p. 3
- RoboPace integrates predicted-contact velocity constraints with robot-limit-aware retiming for online action chunks, leaving the policy and predicted geometry unchanged.The authors state that, to their knowledge, prior work had not integrated these elements in this way.Source: §2.2 · p. 3
3. Problem Formulation
The formulation retimes each policy-predicted chunk by choosing a monotone path parameterization that minimizes execution duration while respecting robot limits and slowing near predicted contact. It separates the chunk’s geometry from its source timing and carries the commanded transition speed into the next chunk.
3. Problem Formulation
- Each chunk is represented by a joint-space path q(s), and the retimer seeks a monotone time parameterization that minimizes chunk duration while slowing near predicted contact.The path is defined over retained waypoints, with contact-critical portions comprising waypoints near predicted or ongoing contact.Source: §3 · p. 4
- The retiming must respect each joint’s velocity and acceleration limits and each hand body’s Cartesian contact-speed limit.The entry path speed corresponds to the commanded joint velocity at the chunk transition, allowing consecutive chunks to join without stopping.Source: §3 · p. 4
- Contact-critical waypoints are obtained from geometric proximity and a learned contact predictor, with the contact-speed limit smoothly ramped around them.The optimization fixes the entry speed and sets the exit speed as specified for contact-aware retiming.Source: §3 · p. 4
4. RoboPace
RoboPace turns each policy chunk into a joint path and retimes it online around predicted contact while preserving its geometry. Its pipeline combines contact cues, robot constraints, and chunk-to-chunk continuity handling.
4.1. Overview
- Each policy chunk becomes a joint path, contact cues mark critical waypoints, and contact-aware TOPP-RA solves its timing online before play-out commands the retimed path.The policy is requeried after a fixed number of commanded source steps; only the contact predictor is learned, while the rest is deterministic given fixed parameters and robot constraints.Source: §4.1 · p. 4
4.2. Action Chunks to Joint Paths
- RoboPace converts each arm’s pose-delta chunk to joint waypoints by sequential inverse kinematics, combines both arms on one path, and drops unreachable or excess waypoints.The combined path compresses short-dwell segments while retaining source indices; forward kinematics supplies hand-body poses for contact cues.Source: §4.2 · p. 4
4.3. Contact Prediction and Gating
- A waypoint is contact-critical when contact is imminent or ongoing, as flagged by either modeled hand-to-hand or hand-to-environment proximity, or wrist-camera prediction.The camera cue estimates source steps to contact from the wrist camera and gripper angle; predictions are aged for observation-to-command and command-to-application delays, and stale predictions are dropped.Source: §4.3 · pp. 4–5
4.4. Contact-Aware Path Retiming
- A gated, position-dependent joint-velocity ceiling translates the contact-speed limit through hand-path rate and replaces the usual joint-velocity ceiling near critical waypoints.The gate ramps toward the critical waypoint. At chunk transitions, commanded velocity is projected onto the new path tangent, while the exit speed is relaxed until feasible.Source: §4.4 · p. 5
- Away from contact, pace follows robot velocity and acceleration limits scaled by σ; with a shared path parameter, either hand’s contact constraint can slow both arms.When hand bodies are stationary along the path, no contact-derived ceiling is imposed, but joint velocity limits still apply.Source: §4.4 · pp. 5–6
4.5. Online Execution and Chunk Transition
- The plan clock resamples motion uniformly in time, while policy requery follows commanded source steps and a wall-clock fallback handles stalls.The gripper chunk is sampled at the source step reached by the arm’s plan clock.Source: §4.5 · p. 6
- Incoming chunks blend with the in-flight path to avoid a step, while rows delayed by pipeline latency are dropped and the remaining plan is shifted to avoid repeating commanded motion.The blend weights the new path against the old path read ahead of the plan clock; the new path stands alone after the old one ends.Source: §4.5 · p. 6
5. Experiments
Experiments compare RoboPace with uniform-rate and physical-limit-only execution across three contact-rich tasks, then examine retiming sensitivity and a bimanual handover.
5.1. Experimental Setup
- The evaluation spans Cup, Socks, and three Pyramid phases across five commands; operator intervention or an unrecoverable task counts as failure.At contact onset, predictor MAE is 0.80 source steps (0.027 s) over 1,540 onsets.Source: §5.1 · pp. 6–8, Figure 4 · p. 7
5.2. Comparison with Baselines
- RoboPace achieves 32 versus 22 successes and takes 45–59% of decelerated successful-trial time on four commands.Accelerated execution succeeds in four of fifty trials, while RACE-style succeeds in one of fifty.Source: Figure 5 · p. 8, §5.2 · pp. 8–9
5.3. Qualitative Analysis of Bimanual Handover
- In one Part handover, RoboPace slows through transfer and succeeds, but the run stops before final placement because held-grasp contact is not predicted.The five quantitative commands contain no hand–hand contact; this qualitative run uses a different transition schedule and requery interval.Source: §5.3 · p. 9, Figure 6 · p. 9
5.4. Retiming Sensitivity
- At 0.02 m/s, four of six schedules achieve four successes in 16–22 s, and Linear-N3 is selected as the default.On Pyramid phase 2, performance is more sensitive to contact speed than transition schedule; lower limits trade speed for reliability, with no consistent schedule ranking.Source: Table 1 · p. 10, §5.4 · p. 10
6. Conclusions and Discussion
RoboPace retimes policy-predicted paths online under contact and robot-motion constraints, achieving better success than slow uniform execution at substantially lower time cost. The evaluation covers one robot and one policy family.
6. Conclusions and Discussion
- RoboPace achieved 64% overall success versus 44% for slow uniform execution, completing four of five commands in about half the slow-execution time.It also exceeded demo-speed success on four of five commands.Source: §6 · p. 11
- RoboPace preserves the policy’s predicted path while imposing contact-dependent velocity ceilings alongside robot kinematic and dynamic constraints, without policy retraining.The method uses a single time-optimal path parameterization to adapt execution timing online.Source: §6 · p. 11
- The formulation applies across policies and robots when multi-step action predictions and target-robot kinematic and dynamic constraints are available, but evaluation covers one robot and one policy family.The authors identify evaluation across robots, policies, tasks, and prediction horizons as future work.Source: §6 · p. 11