ReLoco: Unified Humanoid Locomotion and Fall Recovery via Behavior-Conditioned Motion Priors
Authors: Lequn Fu, Yufeng Hu, Yijun Zhong, Jie Zhang, Xiao Li, Lingxiang Ji, Zhiyuan Xu, Jian Tang, Shiqi Li
Organizations: Huazhong University of Science and Technology, Wuhan 430074, China · Beijing Innovation Center of Humanoid Robots, Beijing 100086, China
Humanoid robots must recover from falls and resume commanded motion to sustain locomotion. Learning these behaviors in one policy requires supervision that accommodates both upright movement and ground-contact recovery. We present ReLoco, a framework that uses behavior-conditioned adversarial motion priors to learn locomotion, get-up from arbitrary fallen postures, and command resumption within one proprioceptive policy. During training, robot posture selects the expert pool and conditions a shared discriminator with two behavior-specific outputs. Task rewards and shared regularization support command tracking and recovery, while temporary assistance facilitates exploration. A simulation-based refinement stage supplies robot-executed motion references for this supervision. Ablations examine conditional supervision, output separation, and reference quality under matched settings. At a common training iteration, the full policy achieves 96.80% independent get-up success and 96.96% push-recovery success, including command resumption. Its torso planar velocity root-mean-square error is 0.1266m/s, 14.7% lower than the conditional single-output variant, with similar joint-angle discrepancy measured by dynamic time warping. Deployment on the 64.5-kg Tiangong3.0 humanoid demonstrates walking, running, get-up, and locomotion resumption after falls using one proprioceptive actor, without controller switching or on-robot policy fine-tuning.
Figures & tables
Fig. 1: Overview of ReLoco. Retargeted human motions are refined in simulation to form locomotion and get-up reference sets. Posture conditions expert sampling and the output of a shared discriminator. Together with task rewards and common regularization, this supervision trains one actor for locomotion and recovery.
(a) Policy observations
Observation term
Actor
Critic
Linear velocity
–
✓
Angular velocity
✓
✓
Projected gravity
✓
✓
Velocity command
✓
✓
Joint positions
✓
✓
TABLE I: Policy observations and training configuration.
Reward Term
Equation
Weight
Upright Task Rewards
Planar Velocity
ρ(vxycmd−Yvxya;st)
1.2
Yaw Velocity
ρ(ωzcmd−ωza;st)
1.2
Hard Velocity Tracking
rtrack,hard
0.5
Torso Orientation
exp[−10(∣ϕa∣+∣θa∣)]
0.8
Prolonged Double Support
χls1[Tds>0.5s]
−100.0
TABLE II: Task-reward terms and shared penalties.
Reward Term
Equation
Weight
Tracking Rewards ( rtrack )
Global Torso Pos.
exp(−∥ptorso−p^torso∥2/σp2)
0.5
Global Torso Ori.
exp(−∥Rtorso⊖R^torso∥2/σR2)
0.5
Relative Body Pos.
exp(−∥pbody−p^body∥2/σb,pos2)
1.0
Relative Body Ori.
exp(−∥Rbody⊖R^body∥2/σb,ori2)
1.0
Body Linear Vel.
exp(−∥vbody−v^body∥2/σv2)
1.0
TABLE III: Reward terms for reference-motion refinement.
Fig. 2: Effect of reference-motion refinement on AMP training. (a) Mean task reward. (b) Get-up success rate. Raw uses the original retargeted motions; Refined uses simulation-refined references. Shading denotes within-run variation over 20,000 training iterations.
Metric
Raw
Refined
Ground penetration (%) ↓
28.99±27.60
0.75±2.09
Self-collision (%) ↓
12.91±14.77
0.35±1.27
Joint-limit violation (%) ↓
12.90±17.24
0.06±0.38
Body-position jerk (m/s 3 ) ↓
406.88±310.44
262.49±118.73
TABLE IV: Reference-motion quality (43 matched motions, mean ± standard deviation).
Fig. 3: Initial recovery assistance. (a) Assistance-force progression. (b) Recorded get-up success. Final rates follow assistance withdrawal; results are from individual training runs.
Method
Torso planar RMSE (m/s) ↓
Joint DTW (rad) ↓
Foot-sliding proxy (m/s) ↓
Get-up SR (%) ↑
Push-recovery SR (%) ↑
I90 (iterations) ↓
Task-only RL
0.0717±0.0006
0.2944±0.0005
0.0635±0.0025
0.00±0.00
8.24±0.92
>20,000
Unconditional AMP
0.8599±0.0554
0.1651±0.0002
0.0956±0.0025
94.64±0.92
92.32±1.25
4,063
Single-head conditional AMP
0.1485±0.0015
0.1303±0.0003
0.0878±0.0019
90.16±2.48
93.36±1.19
6,357
Ours
0.1266±0.0009
0.1307±0.0002
0.0856±0.0019
96.80±1.47
96.96±0.73
3,600
TABLE V: Ablation results at training iteration 20,000 . Metrics are mean ± sample standard deviation over five paired evaluation batches, with 250 trials per batch in each recovery test. Push-recovery SR requires upright recovery and command tracking under the RMSE criterion in Eq. ( 17 ). I90 is the logged training milestone.
Fig. 4: Real-world locomotion on (a) stone-tiled pavement, (b) brick pavement, (c) grass, (d) a downhill grassy slope, (e) an uphill grassy slope, and (f) a treadmill. Standing, walking, and running are generated by the same policy without explicit gait-mode switching; the treadmill belt-speed setting reaches 9.3 km/h.
Method
Success (%) ↑
Time (s) ↓
Mech. work (kJ) ↓
Peak ∣τ∣ (Nm) ↓
Mean ∣q¨∣ (rad/s 2 ) ↓
Reference-motion start pose
HoST [ 34 ]
99.3
1.64±0.45
11.2±2.4
399±4
139±21
Single-ref. tracker [ 43 ]
100
3.84±0.00
1.7±0.0
400±0
17±1
Ours
100
2.32±0.45
2.3±0.6
270±69
36±5
Randomized settled fallen states
HoST
99.9
1.57±0.75
10.7±5.2
400±2
131±21
TABLE VI: Recovery-only comparison under matched initial conditions and success criteria. Success rates use all 1000 episodes per condition; other entries are mean ± standard deviation over successful episodes.
Fig. 5: Autonomous recovery following external kicks from (a) rear, (b) front, and (c) side directions. Yellow dashed boxes highlight representative postural adjustments during the fall-to-recovery transition.
(a) All recorded disturbance attempts
Condition
All attempts
No fall
Fall and recovery
Hardware protection
Standing
47
5
34
8
Walking
10
9
1
0
Running
11
7
3
1
TABLE VII: Hardware disturbance outcomes and recovery timing. (a) All recorded attempts, including protection events; locomotion recoveries resumed the commanded task. No additional task failures were observed. (b) Timing over the first ten successful recoveries per posture (30 trials total), reported as mean ± standard deviation.
Fig. 6: Real-world balance recovery after manual disturbances without a complete ground-level fall. (a) and (b) Responses on approximately level grassy terrain. (c) Response to a push on an inclined grassy surface, showing postural adjustment and restoration of an upright stance. The slope sequence is a qualitative demonstration outside the 20 level-ground locomotion trials.