CALM: Current Aligned Link Manipulation for Single Arm Oversized Object Lifting

I. INTRODUCTION

• Oversized-box lifting is long-horizon, sequential, and contact-rich, while simulated joint effort differs from physical motor current and complicates consistent policy learning across simulation and hardware.Distributed whole-arm tactile sensing generally requires specialized sensorized hardware, and direct hardware reinforcement learning entails extensive interaction, repeated resets, and potentially unsafe exploration.Source: §I · p. 1
• CALM uses a single rigid arm’s intermediate links to form an opposed grasp and lift objects wider than the end-effector aperture, without specialized grasping hardware or distributed whole-arm tactile sensing.The task proceeds through repositioning, opposed grasping, and lifting.Source: §I · pp. 1–2, Figure 1 · p. 1
• CALM maps simulated joint histories to expected physical motor current, then distills routed stage-specific teachers into a deployable student refined with DAgger.The teachers learn repositioning, grasping, and lifting in simulation and generate complete demonstrations under domain randomization.Source: §I · pp. 1–2

A. Whole Arm and Robot Link Manipulation

• Prior whole-arm systems establish link-contact utility but commonly use multiple limbs or specialized compliant and sensing hardware, unlike CALM’s single rigid arm.Earlier approaches include distributed limb contacts –, sensorized compliant platforms,, and other multi-limb or whole-body systems –.Source: A. Whole Arm and Robot Link Manipulation · p. 2

B. Current Alignment for Sim2real Transfer

• CALM causally aligns simulated joint histories with physical motor-current observations, addressing an observation-alignment gap beyond prior work on load representation and actuator modeling.Related methods use motor current for joint-load feedback,, temporal learning, and learned actuator models,.Source: B. Current Alignment for Sim2real… · p. 2

C. Stage Routed Policy Transfer for Long Horizon Tasks

• Because independently trained teachers may not preserve feasible transitions, CALM routes them through handoff states before distilling their behavior into one deployable policy.Prior skill-chaining and multi-teacher approaches include Sequential Dexterity, Robot Parkour, and RLDG; DAgger queries teachers at student-visited states.Source: C. Stage Routed Policy Transfer for… · p. 2

III. METHOD

• The framework maps simulated histories to predicted current, trains privileged stage teachers, routes complete rollouts, distills them into a student, refines with DAgger, and deploys the student on hardware.The overview describes three teachers for repositioning, grasping, and lifting, with privileged simulation state selecting the active teacher during complete rollouts.Source: §III · p. 2, Figure 2 · p. 3

A. Task and Observation Formulation

• The task defines repositioning, opposed wrist-and-upper-arm contact, and lifting; success requires the box bottom to stay at least 0.20 m above the table for 5 seconds.Student training and hardware deployment use sensor observations compatible with deployment, including measured current and camera-estimated box pose in place of simulation-derived channels.Source: A. Task and Observation Formulation · p. 3

B. Causal Current Alignment

• The causal current mapper predicts physical motor current from simulated command, joint-state, and joint-effort histories to align the observation channel before student training.The alignment targets a mismatch caused by simplified actuator, transmission, friction, and low-level control effects in simulation.Source: B. Causal Current Alignment · p. 4
• Paired training data synchronizes commanded joint positions and velocities across simulation and hardware, recording simulated joint effort and physical motor current as respective signals.The commanded motions follow time-parameterized trajectories through sampled Cartesian waypoints.Source: B. Causal Current Alignment · p. 4
• At 120 Hz, the mapper uses a 120-sample, 1-second history to forecast 60 samples, or 0.5 seconds, across six current channels.During student training, the mapper runs at 120 Hz and the 30 Hz student receives the latest mapped current after every four mapper updates.Source: B. Causal Current Alignment · p. 4
• The mapping loss combines Huber error on current magnitudes with penalties on first and second temporal differences.The two difference penalties have corresponding weights.Source: B. Causal Current Alignment · p. 4

C. Privileged Teachers and Router

• The privileged PPO teachers respectively reposition and settle the box, establish the opposed grasp, and lift it, with randomized box dimensions and contact friction among the training variations.Randomized box dimensions are supplied to each teacher so its behavior can adapt to object geometry.Source: C. Privileged Teachers and Router · p. 4
• Table II covers principal rewards and penalties specific to each task stage.The supplied passage identifies the table’s scope but does not provide its reward values.Source: C. Privileged Teachers and Router · p. 4, Table II · p. 5
• Downstream teachers train from successful terminal states produced by the preceding teacher, matching their training states to preceding-stage handoffs.Successful repositioning states reset grasp-teacher training, and successful grasping states reset lift-teacher training.Source: C. Privileged Teachers and Router · p. 5
• A deterministic privileged router selects repositioning, grasping, or lifting according to box and grasp state, with persistence, hysteresis, and minimum dwell times limiting rapid switches.Loss of position or grasp returns control to the corresponding earlier teacher; unsafe or unrecoverable states end the episode.Source: C. Privileged Teachers and Router · p. 5

D. Routed Demonstrations and Domain Randomization

• Complete routed teacher episodes provide imitation samples from repositioning, grasping, and lifting trajectories generated under router control.The recorded student input is the deployable observation history, which remains continuous across teacher switches; the stage label is a supervision target, not a deployed input.Source: D. Routed Demonstrations and Domain… · p. 5
• The imitation dataset retains samples only from complete successful episodes, preserving the routed sequence of repositioning, grasping, and lifting.Episodes are collected under randomization of box dimensions, position and orientation, contact friction, delays, and sensor noise.Source: D. Routed Demonstrations and Domain… · p. 5

E. Stage Aware Student and DAgger

• The student’s shared latent representation feeds an action head and an auxiliary stage head, with stage supervision encouraging the encoder to represent information relevant to each task stage.The predicted stage neither selects a teacher nor chooses an action head at deployment.Source: E. Stage Aware Student and DAgger · p. 5
• The student uses a two-layer GRU of width 256 to process 34-dimensional observations, with action and stage heads sized [256, 256, 6] and [256, 128, 3].Training uses AdamW for 40 epochs, batch size 1,024, and learning rate 3×10−4.Source: E. Stage Aware Student and DAgger · p. 5
• The behavior-cloning objective weights action imitation by stage and adds an auxiliary cross-entropy stage loss.The action term distills the selected teacher’s command; the factor 1/6 averages squared action error across six joints, and λz controls the auxiliary stage loss.Source: E. Stage Aware Student and DAgger · p. 5
• During DAgger rollouts, a Bernoulli selector chooses whether the teacher or student action advances the simulation, while teacher actions and router stage labels are recorded at every visited state.The collected samples are aggregated with the existing dataset to refine the student policy.Source: E. Stage Aware Student and DAgger · p. 5

• Measured joint state, motor current, and camera-estimated box pose replace simulated joint state, predicted current, and simulated box pose.Source: F. Physical Deployment · p. 5
• Measured box dimensions enter through the same normalized task-input channels used in simulation.Source: F. Physical Deployment · p. 5
• Simulation and hardware share channel order, units, normalization, history initialization, observation rate, and action interface; safety limits constrain execution.Deployment uses no teacher, router, privileged stage, exact simulation state, or simulated contact signal.Source: F. Physical Deployment · p. 5

• The first question is whether current alignment reduces the observation gap between simulation and reality.Source: §IV · pp. 5–6, Table III · p. 6
• The second question is whether stage-aware distillation and DAgger improve unified-student performance.Source: §IV · p. 6, Table III · p. 6
• The third question is how reliably the complete CALM system performs physical box lifting.Source: §IV · p. 6, Table III · p. 6

• Both settings use identical six-axis UR7e robots, with 30 Hz incremental position actions tracked by a joint-position PD controller.Simulation and current mapping run at 120 Hz, with each policy action held for four simulation steps.Source: A. Experimental Setup and Evaluation… · p. 6
• Hardware tests use two corrugated-cardboard boxes, with measured dimensions supplied as normalized task inputs and a new random box pose for each trial.Checkpoints, normalization, object properties, camera calibration, and safety limits are fixed before testing.Source: A. Experimental Setup and Evaluation… · p. 6, Figure 3 · p. 6
• A complete-task trial is unsuccessful if it times out or triggers a safety event; simulation also records stage outcomes and unintended contacts.Source: A. Experimental Setup and Evaluation… · p. 6
• Fig. 4 plots joint effort, recorded current, and predictions over 15 seconds, plus joint-wise and mean MAE for 19,213 held-out samples.The simulated-effort baseline uses a separate per-joint linear scaling fitted on training data.Source: Figure 4 · p. 6

B. Current Alignment Accuracy

• Current alignment was assessed on a chronological split, using the first 80% of paired recordings for training and the final 20% for validation.The comparison included a per-joint linear scaling baseline, an MLP, a GRU, and a Transformer, assessed with MAE and RMSE.Source: B. Current Alignment Accuracy · p. 6
• The Transformer achieved 0.04 A aggregate MAE and 0.07 A RMSE, versus 0.06 A MAE and 0.12 A RMSE for the strongest simulation-only baseline, the GRU.The linearly scaled simulated-effort baseline obtained 0.46 A MAE and 0.58 A RMSE across 19,213 held-out samples from one recording session.Source: B. Current Alignment Accuracy · p. 6

C. Teacher Evaluation and Routed Rollouts

• 82.1% of routed episodes completed the task, with conditional handoff success of 99.8% from repositioning to grasping and 99.4% from grasping to lifting.The router can return to grasping to reestablish contact and retry lifting.Source: Figure 5 · p. 7, C. Teacher Evaluation and Routed… · p. 7
• Independent teacher success was 823/1,000 for repositioning, 832/1,000 for grasping, and 142/1,000 for lifting.The routed evaluation sequences the teachers through complete episodes in 1,000 simulation environments.Source: C. Teacher Evaluation and Routed… · p. 7
• All 179 unsuccessful routed episodes timed out: 172 before repositioning, two before grasping, and five after grasping.The authors note that independent lifting's strict grasp-loss criterion counts even small initial box slips, which may be difficult to see visually.Source: C. Teacher Evaluation and Routed… · p. 7

D. Teacher to Student Transfer

• 82.1%, 56.8%, 62.3%, and 76.2% complete-task success were achieved by the routed teacher, action-only BC, stage-aware BC, and DAgger-refined stage-aware student, respectively.Stage supervision added 5.5 percentage points over action-only BC; DAgger added a further 13.9 points, leaving a 5.9-point gap to the routed teacher.Source: D. Teacher to Student Transfer · p. 7

E. Physical CALM Evaluation

• 73.3% (22/30) hardware success was achieved by CALM, compared with 0/10 for the student trained without current alignment.The ablated student had six protective stops, two dropped boxes, and two repositioning timeouts.Source: E. Physical CALM Evaluation · p. 7

V. CONCLUSION AND OUTLOOK

• CALM trains privileged teachers for repositioning, grasping, and lifting, then distills routed demonstrations into a unified student using deployable observations.A Transformer-based current mapper aligns simulated joint effort with physical motor current to provide a loading-sensitive observation channel in simulation and on hardware.Source: §V · p. 7
• The current mapper does not explicitly account for mobile-base motion or other coupled dynamics that alter motor current.The conclusion identifies incorporating these factors as a possible route to more versatile mobile manipulators for physically assistive work.Source: §V · p. 7