During contact-rich manipulation, interactions between a robot and its environment carry information about local geometry: a surface prevents penetration, a bore guides a peg. A controller that exploits these interactions can comply with environmental constraints while preserving task intent. Robot-earning systems commonly use position, hybrid force-position, or Cartesian impedance control. Their prescribed tracking objectives, stiffness, or force-control directions may not match local constraints and may degrade performance. We learn a proprioceptive reflex policy in simulation on three simple interaction primitives: a spring, a plane, and a rail. The policy maps task-space commands to joint-position targets using state history, without direct force or geometrical measurements. Once trained, it serves as a frozen execution layer beneath higher-level controllers. We evaluate it in dual-arm box lifting, peg insertion, and surface following. In box lifting, the reflex kept the force below the threshold while the baseline failed. In rough-surface following it stayed below the 10 N reference, on par with tuned hybrid force-position control. In 0.02 mm peg insertion It reduced mean estimated contact force to less than half that of the baseline while increasing hardware success rates from at most 22 % to 36-58 %. By separating contact response from command generation, the reflex policy provides motion planners, learned policies, and teleoperators with robust contact-rich execution.
Figures & tables
Fig. 1: Overview. (A) Bimanual box lifting on the humanoid (left), and joint-position targets (orange) against the measured robot state (solid) with the reflex (center) and with the nominal IK targets (right). The reflex pulls the targets back toward the box surface, which limits the squeezing force, whereas the nominal targets lie deep inside the box. (B) Peg insertion: the QDD arm (left), training in simulation (center), and the evaluation holes with 0.02 and 0.1 mm tolerances (right). (C) Control hierarchy with the reflex between the task-level command source and the joint-level controller. Bottom: the three interaction primitives used in training: (a) spring, (b) plane, and (c) rail.
Term
Formula
Weight
Pose tracking
Position, coarse
κ(∥xt−xtc∥/0.1)
0.10
Position, fine
κ(∥xt−xtc∥/0.025)
0.05
Orientation
κ(θt/10∘)
0.20
Interaction force
Force tracking
∣ft−ftd∣
−0.02
TABLE I: Reward terms for reflex policy training. The reward is the weighted sum of all terms; negative weights denote penalties. Position errors are measured to the compliant reference xtc (Eq. ( 6 )). Force terms are in newtons and apply only in compliance mode ( mt=1 ).
Fig. 2: Box lifting with the reflex and IK + PD at increasing commanded penetration depths. Left: average joint position tracking error in simulation and on hardware; a larger error means that the joint targets lie farther inside the box (orange in Fig. 1 A). IK + PD diverged on hardware at 17.5 cm. Right: simulated palm contact force; the dotted line marks the 2 N training threshold.
Offset
Success
Time (s)
Est. force (N)
Method
(mm)
count (%)
Mean ± SD
Mean
Peak
0.02 mm tolerance
Reflex
0
29/50 (58)
10.47±4.40
7.41
34.01
2
28/50 (56)
11.22±5.21
7.63
32.72
5
18/50 (36)
11.11±5.37
6.81
40.01
9
21/50 (42)
10.37±4.74
6.33
32.92
TABLE II: Real-world peg insertion with the reflex (ours) or IK + PD as the execution layer beneath a LSTM insertion policy.
Fig. 3: Real-world peg insertion with the reflex. (1–5) Stages of a successful trial. Bottom right: a failed trial, with the peg jammed at the hole entry.
Tolerance (mm)
Controller
0.02 †
0.1 †
1.0
Reflex (Ours)
74.8±4.6
93.1±2.8
99.0±0.3
IK + PD
54.6±12.4
72.8±7.1
97.9±1.6
Impedance
10.4±4.3
39.7±9.9
90.0±2.2
Ours w/o spring
68.2±1.8
65.0±22.7
96.2±0.6
Ours w/o rail
56.9±2.3
76.1±6.8
97.8±0.2
TABLE III: Insertion success rates using different controllers.
Rim-contact time (s)
Rim / all failures (%)
Tol. (mm)
Reflex
IK + PD
Reflex
IK + PD
0.02
2.64
4.14
100/451 (22.2)
323/559 (57.8)
0.1
1.59
2.88
24/123 (19.5)
55/306 (18.0)
TABLE IV: Rim contact and rim failures in simulated peg insertion, with hole-position offsets of 0–3 mm.
Fig. 4: Simulated surface following results. (A) Setup and an example rough surface. (B, C) Flat and rough surfaces: contact force, tangential error, and force p95 versus commanded penetration. Boxes show the median, interquartile range, and 5th–95th percentiles during contact. Dashed lines mark Fth=10 N. (D) Force p95 versus free-space position error; circle size indicates tangential RMSE. Baselines span stiffness scales of 0.25× to 2× ; the reflex is fixed.
How should a robot learn to manipulate objects so fragile that sub-Newton contact forces can cause irreversible damage? Existing visuo-tactile policy learning typically treats tactile sensing as an additional policy input. In direct-contact force-sensitive manipulation, however, the bottleneck can arise earlier, during data collection: manual gripper control is too delayed and coarse-grained to reliably maintain the narrow force range required for stable grasping. We therefore use a deterministic 25 Hz tactile reflex controller as a collection-time teacher, producing demonstrations with controller-shaped grasping behavior for tactile-free policy learning. On Action Chunking with Transformers (ACT), policies trained from reflex-shaped demonstrations recover the teacher's grasping profile and achieve 95% stable grasps on the nominal plastic-cup task, substantially outperforming visually screened manual demonstrations. The same intervention improves in-distribution stability on π0.5 and shows a favorable exploratory trend on an unseen paper-cup variant. Under randomized external disturbance, however, the reflex-data π0.5 policy still fails in 45% of policy-only trials, whereas a deployment-time reflex arbiter retains all grasps. These results reveal a new role for tactile feedback in force-sensitive manipulation: rather than integrating tactile into the policy, we use it as a collection-time teacher that shapes grasping behavior in demonstrations for policy learning, while disturbance rejection remains controller-dependent, revealing the boundary of tactile-free policy.
Ziyan Feng, Zizhao Yuan, Yulong Fu +6
The Hong Kong University of Science and Technology (Guangzhou)
Contact-rich manipulation demands human-like integration of perception and force feedback: vision should guide task progress, while high-frequency interaction control must stabilize contact under uncertainty. Existing learning-based policies often entangle these roles in a monolithic network, trading off global generalization against stable local refinement, while control-centric approaches typically assume a known task structure or learn only controller parameters rather than the structure itself. In this paper, we formalize a physically grounded interaction frame, an instantaneous local basis that decouples force regulation from motion execution, and propose a method to recover it from demonstrations. Based on this, we address both issues by proposing Force Policy, a global-local vision-force policy in which a global policy guides free-space actions using vision, and upon contact, a high-frequency local policy with force feedback estimates the interaction frame and executes hybrid force-position control for stable interaction. Real-world experiments across diverse contact-rich tasks show consistent gains over strong baselines, with more robust contact establishment, more accurate force regulation, and reliable generalization to novel objects with varied geometries and physical properties, ultimately improving both contact stability and execution quality. Project page: https://force-policy.github.io/
Hongjie Fang, Shirun Tang, Mingyu Mei +9
Noematrix · Shanghai Jiao Tong University · Zhejiang University +2
Generative imitation learning policies excel at trajectory planning but lack closed-loop force regulation, while directly incorporating force modalities often requires redesigning or retraining the network. We present UniReflex, a universal plug-and-play framework that equips frozen generative policies with variable impedance control (VIC) for contact regulation, guided by force-direction intent collected during demonstration, without further slow-backbone fine-tuning. By non-invasively intercepting deep latent representations from the action head, UniReflex drives a fast reflex network that decouples active force exertion from external interaction response. This scheme predicts normalized anisotropic stiffness directions for directional compliance allocation. Furthermore, UniReflex integrates an adaptive gating mechanism that enables seamless transitions between position-dominant planning and force-dominant execution. Real-world bimanual experiments demonstrate that UniReflex significantly improves contact stability and success rates while preserving original position accuracy. Our approach achieves 25-66x lower per-step backward latency relative to joint training strategies on the evaluated backbones.
Yan Huang, Shoujie Li, Ziwu Song +1
Tsinghua University Shenzhen International Graduate School · Nanyang Technological University · X Square Robot +1