Humanoid controllers have acquired increasingly diverse motor skills, yet translating these capabilities into precise and stable whole-body manipulation remains challenging. Manipulation tasks are sensitive to execution deviations -- even small errors or unintended steps can cause failure. Introducing Anchor 1, a whole-body controller for precise and stable whole-body manipulation, together with a benchmark spanning diverse coordination demands. Anchor 1 adopts calibrated sparse motion references to preserve task-relevant human motion structure without online joint-space retargeting, leaving kinematic freedom for precise control. To guide this freedom towards stable, physically plausible whole-body coordination, we shape the training distribution through offline curation of robot motions paired with these calibrated sparse references. We further introduce explicit error feedback decoupled from motion encoding, improving control precision while supporting flexible command interfaces. To achieve a systematic evaluation of these capabilities, a benchmark is proposed to assess precision, sustained execution stability, and reachability across diverse whole-body coordination demands. Evaluations in simulation and on hardware show stronger precision and stability across all evaluation tracks. Under the most strict real-world global scoring setting, Anchor 1 achieves a 30% to 80% decrease in tracking error with fewer look-ahead inputs, and without retargeting compared to all existing controllers. Downstream experiments demonstrate more efficient teleoperated data collection and higher task success using the same deployment pipeline as well. We believe Anchor 1 supports precise and stable whole-body control as a practical basis for humanoid task execution.
Figures & tables
Figure 1: Anchor 1– Precise, responsive, and stable whole-body manipulation. (A) Precision. A.1 navigating a narrow passage; A.2 carrying a plate through a narrow passage; A.3 pressing a rice cooker button; and A.4 tracing characters with the end effector. (B) Low-latency response. B.1 accurately striking a fast-moving volleyball; and B.2 rapidly and precisely intercepting a narrow-profile tire. (C) Postural stability. C.1 holding a wooden stool and place it on the floor. C.2 bending to retrieve an orange in a narrow space; and C.3 picking up a water bottle while stably kneeling on one knee. (D) Whole-body coordination. D.1 lifting a bicycle upright; D.2 pulling a grocery cart while placing groceries at designated shelf locations; D.3 bending to arrange bedding; and D.4 steadying a basket with one hand while stacking goods with the other during locomotion.
Figure 2: Benchmark construction and motion coverage. (A) Construction of the In-house, Public, and Reachability tracks, with representative reference sequences and timestamps in seconds; purple markers indicate wrist targets. (B) Joint t-SNE projections of all three tracks, providing a qualitative view of their extensive semantic and pose diversity. (C) Spatial distributions of wrist and ankle positions and pelvis trajectories for the In-house and Public tracks, showing upper and lower-body workspaces, stationary manipulation, and locomotion coverage. The two tracks provide complementary coverage of manipulation-related body configurations and global mobility.
Figure 3: Tracking accuracy and reachability across three tracks. All results are evaluated in MuJoCo. (A, D) Local and global success rates (SR) on Inhouse-track (15,342 motions) and Public-track (2,230 motions). (B, E) Mean position errors in millimeters for horizontal root position (global), wrists, and ankles (local). (C) Global (top) and local (bottom) SR across quasi-static and mobile household task categories in Inhouse-track. Categories include complete motions containing the indicated phases and may overlap. (F) Global (top) and local (bottom) SR by data source in Public-track. (G) Global task SR on 3,210 Reachability-track references, grouped by posture. Left and right variants are pooled for posture-wise SR. (H) Mean slips (left axis) and extra steps (right axis) per Reachability-track motion. Reachability metrics use the evaluation window defined in Sec. 4.3.3 . Superscript G denotes global mode for both Anchor 1 and ScaleBFM. Superscript L denotes re-locate mode for Anchor 1 and reference-forcing mode for ScaleBFM. These mode labels are independent of the local and global scoring frames. SONIC’s G1 superscript identifies its robot encoder; Asym. denotes asymmetric. Metric definitions and aggregation are given in Sec. 4.4.1 .
Control requirements
Wrist
Ankle
Method
Look-ahead (s)
No retargeting
Input
pwG (mm)
rwG ( ∘ )
pwL (mm)
rwL ( ∘ )
paG (mm)
raG ( ∘ )
paL (mm)
raL ( ∘ )
(a) Reachability ( N=51 )
MimicLite [ 35 ]
0.08
✗
R
53.59
5.56
27.87
8.22
50.40
6.58
46.15
6.58
ScaleBFM-14 [ 13 ]
0.10
✗
R
64.18
6.40
46.15
8.26
66.05
8.34
65.75
8.34
ScaleBFM-5 [ 13 ]
0.10
✗
R
72.69
8.34
51.22
8.25
77.13
9.68
74.64
9.68
SONIC-G1 [ 12 ]
0.90
✗
R
98.15
10.19
42.38
9.42
92.46
9.45
69.49
9.45
Table 1: Real-world tracking errors and reference requirements across subsets of two tracks. Reachability comprises 51 standing reaches selected to broadly cover the workspace; In-house comprises 72 tag-selected whole-body manipulation motions. Both subsets report global and local wrist and ankle position and orientation errors. Reachability uses the right wrist and bilateral ankle means; in-house uses bilateral means for both wrists and ankles. p and r denote mean position error (mm) and mean orientation error (degrees), respectively. Subscripts w and a identify wrists and ankles; superscripts G and L denote global and local scoring frames, respectively. N denotes the target motion count for each subset. Look-ahead is the maximum future-reference horizon in seconds, excluding historical inputs. Input identifies the reference format ( R , H , or H5′ ). The control requirements apply to both subsets. ✓ indicates that online retargeting is not required; ✗ indicates that it is required. Lower errors are better; bold values indicate the lowest error in each metric within each subset.
Figure 4: Real-world task success across position thresholds on a curated subset of the In-house track. Radar plots show local success rates across seven task categories for 72 motions selected by predefined action tags for real-robot testing. Panels A–F vary the wrist and ankle position tolerance from 10 to 25 cm, with all other success thresholds matching simulation ( Eq. 2 ). Relaxing the tolerance narrows several category-wise differences, while Anchor 1 maintains the highest overall success rate.
Figure 5: Wrist reference fidelity before motion encoding. (A) Right-wrist trajectories for a circular motion with a 16-cm human radius. Gray dashed, green, and orange curves denote the human reference, H5′ targets, and SMPL forward-kinematics (FK) targets, respectively. Shoulder-relative positions are expressed in common human chest axes and normalized by each representation’s fixed arm length. The H5′ trajectory more closely follows the circular shape of the human reference than the SMPL FK trajectory. (B) Unit-PDV scores over time for the circular motion in (A); H5′ yields an 80.2% lower mean score than the SMPL encoder input, indicating better preservation of human wrist motion directions. (C) Schematic illustration of the unit-PDV calculation using time-matched displacement vectors to compare the human reference with H5′ (left) and the SMPL encoder input (right). Calculation details are provided in Sec. S2 .
Figure 6: Motion intent preservation and control stability under retargeting failure. All configurations use the same offline human motion trajectory recorded with Pico, with time progressing from left to right. (A) Human reference motion visualized in SMPL-X format. (B) Anchor 1 with direct sparse H5′ conditioning. (C) SONIC in G1 encoder mode with robot references generated by GMR, an inverse-kinematics (IK) retargeter. (D) SONIC in G1 encoder mode with robot references generated by UMR, a learned point-cloud-matching retargeter. During upper-limb extension, the retargeting-dependent configurations exhibit deviations from the intended arm motion and abrupt changes, whereas Anchor 1 retains the reference motion with stable behavior. Green and red outlines highlight motion preservation and retargeting-related failures, respectively.
Variant
Feedback
Inference
Feedback
Global
Local
location
mode
source
SR (%) ↑
Pos. err. (mm) ↓
SR (%) ↑
Pos. err. (mm) ↓
w/o explicit error feedback
None
Global
—
70.9
28.4
61.7
24.5
Anchor 1 ∗
Encoder
Global
Ref.
62.7
35.6
86.0
20.2
Anchor 1
Decoder
Global
Ref.
88.7
25.5
88.4
21.9
Anchor 1
Decoder
Global
Recon.
76.0
27.7
77.5
22.4
Anchor 1
Decoder
Re-locate
Ref.
35.5
57.9
85.0
22.3
Table 2: Effects of explicit feedback, inference mode, and target reconstruction. All variants use H5′ as input and are evaluated in MuJoCo on the In-house track. Anchor 1 ∗ supplies explicit errors to the encoder; Anchor 1 supplies them to the decoder. Ref. and Recon. denote supplied and reconstructed feedback targets. SR and Pos. err. denote success rate and mean position error over 14 body points. Global and local scoring frames are independent of inference mode.
Figure 7: Motion token interpolation with coupled and decoupled feedback. Top: reference poses (gray) and robot motion under feedback-coupled (orange) and feedback-decoupled (blue) token interpolation at matching times. Endpoints A and B correspond to 0.4 and 2.0 s, respectively. The feedback-decoupled variant more closely follows the transition from a crouched posture toward the upright endpoint. Bottom: root mean square error (RMSE) across 29 joint angles relative to the reference, in degrees. The shaded region marks the interpolation interval; the feedback-decoupled variant has lower error in its later portion.
Figure 8: Precise and stable teleoperation across operators and input devices. (A) Comparison with the baseline during lateral reaching toward the ankle, small-amplitude lateral translation, ankle rotation, and bending forward to reach the ground. (B) Adaptation to operators with heights of 165–188 cm and body masses of 50–80 kg while maintaining stable control and preserving human motion intent. Labels show height (cm) / body mass (kg). (C) Support for different motion-capture devices by computing H5′ directly, without retargeting. From left to right: Xsens (inertial motion capture), CHINGMU (optical motion capture), Noitom Perception Neuron (PN) Link (wired inertial motion capture), and PICO (virtual reality, VR).
Controller
Task
Successful demos/h ↑
Collection SR (%) ↑
SONIC v1.1 (SMPL)
Kick cardboard box
9
37.50
Pick up slippers
41
68.33
Anchor 1 ∗
Kick cardboard box
45
94.74
Pick up slippers
67
93.10
Anchor 1
Kick cardboard box
43
98.03
Pick up slippers
62
97.37
Table 3: Teleoperation data collection efficiency. Successful demonstrations collected per hour and collection success rates for each task. Anchor 1 ∗ denotes the variant with error feedback supplied to the encoder. Higher values are better.
Controller (mode)
Interface
Kick box
Pick up slippers
SR
Score
SR
Score
SONIC v1.1 (SMPL)
Motion token
0.70±0.14
0.63±0.08
0.10±0.00
0.44±0.05
Anchor 1 ∗ (re-locate)
Sparse
0.85±0.07
0.79±0.05
0.45±0.07
0.74±0.05
Anchor 1 ∗ (re-locate)
Motion token
0.00±0.00
0.00±0.00
0.00±0.00
0.00±0.00
Anchor 1 (re-locate + Recon. feedback)
Sparse
0.75±0.07
0.56±0.03
0.20±0.14
0.68±0.04
Anchor 1 (re-locate + Recon. feedback)
Motion token
0.90±0.00
0.81±0.02
0.45±0.07
0.64±0.05
Table 4: Downstream task performance by inference controller and policy interface. SR denotes success rate. Score is weighted by the number of kicks for Kick box and averaged over evaluation steps for Pick up slippers. Both metrics range from 0 to 1. Anchor 1 ∗ uses encoder error feedback. Best results are bolded, including ties.
Figure 9: Autonomous execution of downstream whole-body manipulation tasks. The robot autonomously performs two tasks: (A) kicking a cardboard box into a marked target area and (B) retrieving slippers from a low shelf and placing them at a designated location. Each row shows the progression from approach through object interaction to task completion, with time advancing from left to right.
Figure 10: Overview of Anchor 1. (A) Human motion H is calibrated to obtain H′ and retargeted offline to obtain robot references R ; quality control retains matched pairs and shapes the training distribution. (B) The motion encoder produces tokens for control and target reconstruction. A target selector chooses supplied or reconstructed poses for explicit error feedback, which enters the control decoder alongside robot state. Future prediction is used for training only. (C) Teleoperation supplies sparse motion references; higher-level policies can supply sparse references (Mode 1) or motion tokens (Mode 2). The corresponding feedback targets are supplied references ( H′ ) or reconstructed poses ( H∗ ).
Source
Sequences
Original duration (h)
Mirrored duration (h)
Total duration (h)
Duration share (%)
In-house
288,215
496.04
339.62
835.67
77.28
Chingmu
222,400
227.97
217.19
445.16
41.17
Nokov
7,149
13.67
15.61
29.27
2.71
PN-LINK (locomotion)
58,666
254.41
106.83
361.24
33.41
Public
118,007
192.33
53.37
245.70
22.72
BONES-SEED
74,371
138.29
0.00
138.29
12.79
Table 5: Composition of the training dataset. Sequence counts include original and mirrored sequences. Duration percentages are calculated relative to the total duration after mirroring. Totals are computed before rounding.
Upper body
U1: Torso configuration
U1-1
Upright torso
U1-2
Forward bending
U1-3
Lateral bending
U1-4
Torso rotation
U1-5
Forward bending with rotation
Table 6: First-level and second-level reference tags for coverage analysis. U and L denote upper-body and lower-body descriptors, respectively; 1 and 2 indicate levels of description.
Figure 11: Task construction and reference motion for Mobile-151. (A) upper-body and lower-body action descriptions and task attributes define the task. (B) Representative reference poses, with timestamps in seconds.
Source
Selection protocol
Sequences
BONES-SEED [ 30 ]
Held-out set of 18 actors
1,214
CIRCLE [ 31 ]
Held-out subject s5
442
OMOMO [ 32 ]
Existing test partition
474
AMASS [ 33 ]
Predefined local test subset
100
Total
2,230
Table 7: Sources and selection protocols for the Public track.
Parameter
Evaluation setting
Random seed
0; deterministic sampling by motion and parameter
Robot static friction
U[0.3,1.6]
Robot dynamic friction
U[0.3,1.2]
Robot restitution
U[0,0.5]
Contact-material assignment
64 shared material buckets, assigned per collision shape
Ground material
Friction 1; restitution 0; fixed
Table 8: Domain randomization and sampling settings for simulation evaluation. DR denotes domain randomization and CoM denotes center of mass. U[a,b] denotes uniform sampling on [a,b] . Randomized physical parameters are applied before each motion and remain fixed throughout its rollout.
Track/source
Tags
Upper-body pairs
Lower-body pairs
All pairs
In-house
39/43 (90.70%)
30/33 (90.91%)
18/20 (90.00%)
48/53 (90.57%)
BONES-SEED
28/43 (65.12%)
9/33 (27.27%)
11/20 (55.00%)
20/53 (37.74%)
CIRCLE
22/43 (51.16%)
5/33 (15.15%)
10/20 (50.00%)
15/53 (28.30%)
OMOMO
25/43 (58.14%)
6/33 (18.18%)
11/20 (55.00%)
17/53 (32.08%)
AMASS
26/43 (60.47%)
4/33 (12.12%)
11/20 (55.00%)
15/53 (28.30%)
Public (combined)
29/43 (67.44%)
11/33 (33.33%)
11/20 (55.00%)
22/53 (41.51%)
Table S1: Action-tag and combination coverage across tracks and public data sources. Entries report covered/reference counts and percentages. Tag and pair denominators are fixed across all evaluated sets. Upper-body and lower-body pairs link first-level and second-level tags within the respective body description; all pairs combine these two sets. Public (combined) reports the union of distinct tags and compatible pairs across the four public sources.
Figure S1: Prevalence of second-level action tags. (A) Arm and hand behavior. (B) Stances and movement patterns. Tag codes follow Table 6 . Prevalence is the number of sequences assigned a tag divided by the total number of sequences in the respective track: 15,342 for the In-house track and 2,230 for the Public track. Each full sequence is counted once per tag and may carry multiple tags, so percentages need not sum to 100%. Estimates for the Public track use filename-based candidate annotations; NR denotes unresolved tags in this track, which are not plotted as zero. Lines connect discrete categories as visual guides.